<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>JavaRubberDuck</title>
  <subtitle>A developer&#39;s blog about Java, Cloud, and everything in between. Tips, tutorials, and thoughts for developers.</subtitle>
  <link href="https://javarubberduck.com/feed.xml" rel="self"/>
  <link href="https://javarubberduck.com/"/>
  <updated>2026-11-08T00:00:00.000Z</updated>
  <id>https://javarubberduck.com/</id>
  <author>
    <name>Stanislav Lentsov</name>
    <email>toenail-resolve.0q@icloud.com</email>
  </author>
  
  <entry>
    <title>Spring Ecosystem Update: Spring Boot 4.2.0-RC2, Spring Cloud Paddington RC2 Adds GCS Config Backend, Spring AI MCP Apps Recipes, 2026-11-08</title>
    <link href="https://javarubberduck.com/java/news-2026-11-08-spring/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-11-08-spring/</id>
    <content type="html">&lt;h2 id=&quot;spring-boot-4.2.0-rc2-available-now&quot; tabindex=&quot;-1&quot;&gt;Spring Boot 4.2.0-RC2 Available Now&lt;/h2&gt;
&lt;p&gt;Spring Boot 4.2.0-RC2 is on Maven Central, two weeks after RC1. The release contains 99 enhancements, documentation improvements, dependency upgrades and bug fixes. It is the last planned stop before the 4.2.0 GA later this month.&lt;/p&gt;
&lt;p&gt;Two changes stand out. First, Reactor Netty-based WebFlux applications now honor the &lt;code&gt;X-Forwarded-Prefix&lt;/code&gt; header, which fixes generated links and redirects for apps deployed behind a path-rewriting reverse proxy or ingress. Second, the &lt;code&gt;spring.data.redis.database&lt;/code&gt; property now applies to master/replica connections as well. Previously the setting was ignored there, so replica reads could silently hit database 0.&lt;/p&gt;
&lt;p&gt;Because Spring Boot 4.0 OSS support ends on December 31, 2026, teams still on 4.0 should test against RC2 now so that a later GA upgrade is a formality. The full release notes are on the Spring Boot wiki.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/06/spring-boot-4-2-0-RC2-available-now&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;spring-cloud-2026.0.0-rc2-(paddington)-released&quot; tabindex=&quot;-1&quot;&gt;Spring Cloud 2026.0.0-RC2 (Paddington) Released&lt;/h2&gt;
&lt;p&gt;The Spring Cloud 2026.0.0-RC2 release train, codenamed Paddington, is built on Spring Boot 4.2.0-RC2. Every module in the train is at 5.1.0-RC2. Artifacts are published to the Spring milestone repository.&lt;/p&gt;
&lt;p&gt;Notable additions in this RC:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Spring Cloud Config&lt;/strong&gt; adds Google Cloud Storage as a Config Server backend and can serve plain-text and binary files as a &lt;code&gt;Resource&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Cloud Gateway&lt;/strong&gt; adds a &lt;code&gt;LocalResponseCache&lt;/code&gt; filter for the Server Web MVC flavor, matching what WebFlux users already had.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Cloud OpenFeign&lt;/strong&gt; now sends nested &lt;code&gt;@SpringQueryMap&lt;/code&gt; objects as nested query parameters instead of flattening or dropping them.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The GCS backend is the most useful of these for teams on Google Cloud, who no longer need a Git repo or a custom &lt;code&gt;EnvironmentRepository&lt;/code&gt; to keep config next to their other GCP resources. As announced with RC1, Spring Cloud Bus is now part of Spring Cloud Stream, so check your dependency coordinates when you upgrade.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/05/spring-cloud-2026-0-0-RC2-has-been-released&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;this-week-in-spring%3A-cloud-foundry-in-kind-and-spring-ai-recipes-for-mcp-apps&quot; tabindex=&quot;-1&quot;&gt;This Week in Spring: Cloud Foundry in kind and Spring AI Recipes for MCP Apps&lt;/h2&gt;
&lt;p&gt;Josh Long&#39;s November 3 &amp;quot;This Week in Spring&amp;quot; is mostly a hands-on tutorial. It covers running Cloud Foundry locally inside a kind (Kubernetes-in-Docker) cluster, installing the &lt;code&gt;cf&lt;/code&gt; CLI, and pushing a Spring Boot app, which gives a disposable local CF environment without a full platform install.&lt;/p&gt;
&lt;p&gt;The roundup also links new Spring AI recipes from Craig Walls on two topics: rendering rich UIs through MCP Apps, the interactive UI extension in the 2026-07-28 MCP specification, and speeding up RAG pipelines with faster embedding generation. There is also a step-by-step Spring Boot 3.5 to 4 migration guide, which is timely now that 3.5 OSS support has ended, and a pointer to coverage of the Spring Boot 4.2 release candidate.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/03/this-week-in-spring-november-3rd-2026&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/06/spring-boot-4-2-0-RC2-available-now&quot;&gt;Spring Boot 4.2.0-RC2 available now&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/05/spring-cloud-2026-0-0-RC2-has-been-released&quot;&gt;Spring Cloud 2026.0.0-RC2 (aka Paddington) Has Been Released&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/11/03/this-week-in-spring-november-3rd-2026&quot;&gt;This Week in Spring - November 3rd, 2026&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Spring Boot 4.2.0-RC2 shipped with 99 changes ahead of the November GA, and the Spring Cloud 2026.0.0-RC2 release train followed on top of it.</summary>
    
  </entry>
  
  <entry>
    <title>JavaRubberDuck Podcast — November 2–8, 2026</title>
    <link href="https://javarubberduck.com/podcast/2026-11-08/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/podcast/2026-11-08/</id>
    <content type="html">&lt;h2 id=&quot;in-this-episode&quot; tabindex=&quot;-1&quot;&gt;In This Episode&lt;/h2&gt;
&lt;p&gt;Stories covered from 2026-11-02 to 2026-11-08:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;GitHub Agentic Workflows Reach GA with a Defense-in-Depth Execution Model&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Anthropic Moves Its Monorepo from Git to Scion to Keep Up with Agent Commit Volume&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/anthropic-source-control/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;IBM Granite Switch 4.1.1 Treats Model Capabilities as Versioned, Swappable Adapters&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/ibm-granite/granite-switch-4-1-1&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Docker&#39;s &amp;quot;The CVE That Did Not Exist&amp;quot;: Sandboxes Contain Actions, Not Claims&lt;/strong&gt; — Docker (&lt;a href=&quot;https://www.docker.com/blog/coding-agent-horror-stories-the-cve-that-did-not-exist/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Google&#39;s PageBreak Only Reports Vulnerabilities a Deterministic Validator Can Reproduce&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/google-pagebreak-security-agent/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Gemini Enterprise Semantic Governance Now Evaluates Tool Calls from Non-Gemini Models&lt;/strong&gt; — Google Cloud (&lt;a href=&quot;https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Copilot Fusion Routes Each Prompt Across Multiple Models&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Copilot Comes Bundled with GitHub Enterprise, and the Copilot SDK Moves to Microsoft&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-enterprise-now-includes-access-to-github-copilot&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Agent Merge and Shared Copilot Automations Push Copilot Further into the Cloud&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-28-build-shared-workflows-for-your-team-with-copilot-automations&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GPT-6.1 Terra Arrives in Copilot and Codex, Codex CLI Adds a Vertex AI Provider&lt;/strong&gt; — OpenAI (&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Claude Code 2.1.311–2.1.316: Write Boundaries, Subagent Effort and Gateway Budgets&lt;/strong&gt; — Anthropic (&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Mistral Large 4 Released as a 1T-Parameter Open-Weight MoE&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/mistral-large4&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Microsoft 365 Copilot Adds Admin-Built MCP Connectors and Usage-Billed Work IQ APIs&lt;/strong&gt; — Microsoft Learn (&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Anthropic Files Public S-1 for a Nasdaq IPO Under the Ticker ANT&lt;/strong&gt; — Anthropic (&lt;a href=&quot;https://www.anthropic.com/news/s1-sec&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hugging Face Proposes Model Integration Contracts for Transformers&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/transformers-mic&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CWE-bench v2 Measures Whether Models Can Fix Real Vulnerability Classes&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/cwe-bench&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitLab 19.5 Makes External Coding Agents a Governed Part of the Platform&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/11/gitlab-19-5-agentic-platform/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Where Should an Agent&#39;s Code Live? Docker Compares Mounts, Worktrees and Clone Mode&lt;/strong&gt; — Docker Blog (&lt;a href=&quot;https://www.docker.com/blog/docker-sandbox-code-checkout-workflows/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Docker Sandboxes vs. Containers: Why Unattended Agents Need a Separate Kernel&lt;/strong&gt; — Docker Blog (&lt;a href=&quot;https://www.docker.com/blog/docker-sandboxes-vs-containers/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Edera&#39;s Object Capability Bounded Authority Keeps Credentials Out of Agent Workloads&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/11/ai-credentials-isolation/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Actions Lockfiles Reach GA, but Do Not Make Dependencies Safe&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/11/github-actions-lockfiles-ga/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Copilot Arrives in Xcode&#39;s Coding Assistant and JetBrains Air&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-11-03-github-copilot-is-now-available-in-jetbrains-air&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;@GitHub in Slack and Microsoft Teams Is Generally Available&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-11-06-github-copilot-in-slack-and-teams-is-generally-available&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Claude Code 2.1.317–2.1.319: Background MCP Tasks and Deny-Instead-of-Prompt&lt;/strong&gt; — Claude Code Changelog (&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Codex CLI 0.167.0 Adds an Agent Command Center&lt;/strong&gt; — OpenAI Codex (GitHub) (&lt;a href=&quot;https://github.com/openai/codex/releases/tag/rust-v0.167.0&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Microsoft 365 Copilot: Cowork Drives Your Local Edge, Admins Get an Agents Usage Report&lt;/strong&gt; — Microsoft Learn (&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hermes Agent v0.23.0 &amp;quot;The Hearth Release&amp;quot; Runs the Desktop Workspace in a Browser&lt;/strong&gt; — NousResearch (GitHub) (&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.0&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Google DeepMind Releases Cheminformatics Skills for Antigravity&lt;/strong&gt; — Google DeepMind (&lt;a href=&quot;https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;listen&quot; tabindex=&quot;-1&quot;&gt;Listen&lt;/h2&gt;
&lt;audio controls=&quot;&quot;&gt;
  &lt;source src=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-11-08-podcast.mp3&quot; type=&quot;audio/mpeg&quot; /&gt;
  &lt;a href=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-11-08-podcast.mp3&quot;&gt;Download episode&lt;/a&gt;
&lt;/audio&gt;
&lt;blockquote&gt;
&lt;p&gt;🤖 &lt;strong&gt;AI-generated episode.&lt;/strong&gt; Both hosts are synthetic voices produced with Google NotebookLM — no human narrators were recorded. The conversation is generated from the news digest articles linked above.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;subscribe&quot; tabindex=&quot;-1&quot;&gt;Subscribe&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://javarubberduck.com/podcast.xml&quot;&gt;RSS feed&lt;/a&gt; — paste into any podcast app&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://open.spotify.com/show/&quot;&gt;Spotify&lt;/a&gt; — coming soon&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>GitHub Agentic Workflows go GA, Docker argues unattended agents need microVM sandboxes, and Copilot spreads into Xcode, JetBrains Air, Slack and Teams.</summary>
    
  </entry>
  
  <entry>
    <title>Java News: JDK 28 Closes In on Rampdown, TornadoVM 8.0 Drops Graal, Project Detroit Embeds V8 and CPython, 2026-11-08</title>
    <link href="https://javarubberduck.com/java/news-2026-11-08-java-jdk/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-11-08-java-jdk/</id>
    <content type="html">&lt;h2 id=&quot;jdk-28-nears-rampdown-phase-one-with-most-jeps-integrated&quot; tabindex=&quot;-1&quot;&gt;JDK 28 Nears Rampdown Phase One with Most JEPs Integrated&lt;/h2&gt;
&lt;p&gt;The JDK 28 project page (last updated November 2) confirms the release schedule: Rampdown Phase One on December 3, 2026, when the feature set is frozen and the stabilization branch is cut. Rampdown Phase Two follows on January 14, 2027, the Release Candidate phase starts February 4, and General Availability is set for March 9, 2027.&lt;/p&gt;
&lt;p&gt;With less than four weeks until the freeze, almost all of the ten targeted JEPs have moved past Targeted. The OpenJDK JEP index now lists these as &lt;strong&gt;Integrated&lt;/strong&gt; in the mainline: JEP 401 (Value Objects, Preview), JEP 539 (Strict Field Initialization in the JVM, Preview), JEP 540 (Simple JSON API, Incubator), JEP 543 (Structured Concurrency, final), JEP 544 (Ahead-of-Time Code Compilation), JEP 545 (Faster Startup and Warmup with ZGC), and JEP 546 (Adaptive Heap Sizing for ZGC). JEP 542 (PEM Encodings of Cryptographic Objects) is Completed and JEP 541 (Deprecate the macOS/x64 Port for Removal) is delivered. JEP 535 (Shenandoah GC generational mode by default) is the only one still at Targeted.&lt;/p&gt;
&lt;p&gt;Two Candidates have JDK 28 in view but are not yet targeted: JEP 528 (Post-Mortem Crash Analysis with jcmd) and JEP 547 (Lazy Constants, Fourth Preview). Anything not proposed and targeted in the next few weeks will slip to JDK 29. Value classes and AOT code compilation are already in the early-access builds at &lt;a href=&quot;http://jdk.java.net/28&quot;&gt;jdk.java.net/28&lt;/a&gt;, so now is a good time to run test suites against them. Value classes in particular change identity semantics for migrated JDK types.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://openjdk.org/projects/jdk/28/&quot;&gt;Read more&lt;/a&gt; — OpenJDK&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;java-news-roundup%3A-tornadovm-8.0-ships-its-own-compiler%2C-langchain4j-1.23%2C-hibernate-orm-8.0-beta4&quot; tabindex=&quot;-1&quot;&gt;Java News Roundup: TornadoVM 8.0 Ships Its Own Compiler, LangChain4j 1.23, Hibernate ORM 8.0 Beta4&lt;/h2&gt;
&lt;p&gt;InfoQ&#39;s roundup for the week of October 26 covers JDK 28 early-access Build 22 and the GA release of &lt;strong&gt;TornadoVM 8.0.0&lt;/strong&gt;. The headline change in TornadoVM 8 is a pure-Java compiler, the TornadoVM Compiler (TC), which replaces its previous dependency on the Graal/JVMCI path for generating GPU and accelerator code. That makes TornadoVM much easier to run on stock JDK distributions. The release also uses JEP 516 (Ahead-of-Time Object Caching with Any GC) to cut startup overhead.&lt;/p&gt;
&lt;p&gt;On the framework side, LangChain4j 1.23.0 shipped as a formal release alongside its thirty-third beta. Hibernate ORM reached 8.0.0.Beta4, Hibernate Reactive shipped 5.1.0.Beta1 and 5.0.0, and JobRunr got a 9.0.3 point release. Tooling updates include Eclipse JNoSQL 1.2.0, Endive 1.2.0, JBang 0.143.0, JReleaser 1.27.0 and the first release candidate of Gradle 9.9.0.&lt;/p&gt;
&lt;p&gt;The roundup also introduces the Quarkus Almanac, a new reference resource from the Quarkus team. Anyone tracking the Hibernate 8 migration should read the Beta4 notes now, because the ORM 8.0 line is approaching feature-complete.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/java-news-roundup-oct26-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;project-detroit%3A-embedding-v8-and-cpython-in-java-via-the-ffm-api&quot; tabindex=&quot;-1&quot;&gt;Project Detroit: Embedding V8 and CPython in Java via the FFM API&lt;/h2&gt;
&lt;p&gt;In a Devoxx Belgium 2026 session published on Inside Java, Maurizio Cimadamore presented Project Detroit, which takes a different approach to running other languages on the JVM. Rather than reimplementing a language runtime in Java, as Nashorn did for JavaScript, Detroit uses the Foreign Function &amp;amp; Memory (FFM) API to embed existing native runtimes such as V8 and CPython directly in a Java application.&lt;/p&gt;
&lt;p&gt;The talk contrasts this with Nashorn&#39;s history: shipped in JDK 8, deprecated in Java 11 and removed in Java 15, largely because tracking the evolving ECMAScript spec in a separate Java implementation was too costly. Embedding the reference runtimes gives full language compatibility from day one. The hard engineering problems move elsewhere: two garbage collectors have to cooperate when objects cross the boundary, and calls into native code must not pin virtual threads.&lt;/p&gt;
&lt;p&gt;This matters for teams that today shell out to Node or Python processes for scripting, data science libraries or server-side rendering. An FFM-based embedding could make those in-process calls with JVM-managed lifecycles. Detroit is still an OpenJDK project rather than a JEP, so there is no release target yet.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://inside.java/2026/11/04/devoxxbelgium-project-detroit/&quot;&gt;Read more&lt;/a&gt; — Inside Java&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;inside-java-podcast%3A-%22ask-the-architects%22-at-devoxx-belgium&quot; tabindex=&quot;-1&quot;&gt;Inside Java Podcast: &amp;quot;Ask the Architects&amp;quot; at Devoxx Belgium&lt;/h2&gt;
&lt;p&gt;Episode 74 of the Inside Java Podcast is a recording of the &amp;quot;Ask the Architects&amp;quot; session at Devoxx Belgium 2026. The panel included Georges Saab, Alan Bateman, Ron Pressler, Gavin Bierman, Maurizio Cimadamore, Viktor Klang, David Simms, Per-Åke Minborg and host Nicolai Parlog, and it answered audience questions.&lt;/p&gt;
&lt;p&gt;Topics covered a custom virtual thread scheduler, the OpenJDK approach to deprecations and removals, Javadoc and AI, how OpenJDK sets priorities in the age of AI, primitive specialization for generics after value classes, better modules, memory layout introspection, the future of string templates, easier object construction, and automatic heap sizing. The show notes link Ron Pressler&#39;s Devoxx keynote, &amp;quot;Java - The Year of the Runtime!&amp;quot;.&lt;/p&gt;
&lt;p&gt;No new JEPs were announced, but the session is a useful signal of where the architects are heading after Valhalla&#39;s first preview. Primitive specialization and construction ergonomics both came up as next steps.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://inside.java/2026/11/05/podcast-074/&quot;&gt;Read more&lt;/a&gt; — Inside Java&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://openjdk.org/projects/jdk/28/&quot;&gt;JDK 28 Project Page&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/java-news-roundup-oct26-2026/&quot;&gt;Java News Roundup: TornadoVM 8.0, JNoSQL, Endive, JReleaser, JBang, LangChain4j, Quarkus Almanac&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://inside.java/2026/11/04/devoxxbelgium-project-detroit/&quot;&gt;Project Detroit: From Foreign Libraries to Foreign Language Runtimes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://inside.java/2026/11/05/podcast-074/&quot;&gt;Episode 74 &amp;quot;Ask the Architects at Devoxx&amp;quot;&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>JDK 28 heads into Rampdown Phase One on December 3 with most of its ten JEPs already integrated, including Value Objects (Preview), AOT code compilation and Structured Concurrency.</summary>
    
  </entry>
  
  <entry>
    <title>Developer Tools Digest: Copilot Lands in Xcode and JetBrains Air, Claude Code Backgrounds MCP Tasks, Hermes Agent Goes to the Browser, 2026-11-08</title>
    <link href="https://javarubberduck.com/ai/news-2026-11-08-dev-tools/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-11-08-dev-tools/</id>
    <content type="html">&lt;h2 id=&quot;github-copilot-arrives-in-xcode&#39;s-coding-assistant-and-jetbrains-air&quot; tabindex=&quot;-1&quot;&gt;GitHub Copilot Arrives in Xcode&#39;s Coding Assistant and JetBrains Air&lt;/h2&gt;
&lt;p&gt;GitHub shipped two integrations that put Copilot inside other vendors&#39; agent surfaces instead of its own extensions. On November 4, Copilot became available in Xcode&#39;s built-in coding assistant in public preview, starting with Xcode 27.3 beta on macOS Tahoe 26.6 or later. It is set up under Xcode &amp;gt; Settings &amp;gt; Intelligence and works with every Copilot plan. It sits next to the existing GitHub Copilot for Xcode extension rather than replacing it.&lt;/p&gt;
&lt;p&gt;A day earlier, Copilot became an agent inside JetBrains Air via the Agent Client Protocol (ACP). It supports local tasks in a JetBrains IDE and cloud tasks through Air Teams, which run in a Docker container on a fresh VM. Both are billed to the user&#39;s Copilot subscription rather than JetBrains AI credits. Air administrators can disable Copilot by governance policy, but they cannot set Copilot spend limits from Air.&lt;/p&gt;
&lt;p&gt;The JetBrains integration is another sign that ACP is becoming the standard socket for coding agents in editors, the way LSP did for language tooling. GitHub&#39;s weekly release notes for the same week also list VS Code 1.145 (chat view modes, remote browser preview, cron expressions for local automations), a unified &lt;code&gt;/settings&lt;/code&gt; dashboard in Copilot CLI, and Agent Merge setup directly from a pull request in the Copilot app.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-03-github-copilot-is-now-available-in-jetbrains-air&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;%40github-in-slack-and-microsoft-teams-is-generally-available&quot; tabindex=&quot;-1&quot;&gt;@GitHub in Slack and Microsoft Teams Is Generally Available&lt;/h2&gt;
&lt;p&gt;The &lt;code&gt;@GitHub&lt;/code&gt; bot for Slack and Microsoft Teams reached general availability on November 6. From a conversation, it can investigate bugs, create issues, implement changes and open pull requests, using the thread as context so a team discussion can turn into a tracked change without switching tools.&lt;/p&gt;
&lt;p&gt;The GA release adds Canvas support in Teams for reviewing plans and code inline, automatic repository detection from the message content, and attribution on created issues and PRs (&amp;quot;Created from a Slack/Teams conversation with @username&amp;quot;). The attribution helps audits, because agent-authored PRs can be traced back to the conversation and person that requested them.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-06-github-copilot-in-slack-and-teams-is-generally-available&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;claude-code-2.1.317%E2%80%932.1.319%3A-background-mcp-tasks-and-deny-instead-of-prompt&quot; tabindex=&quot;-1&quot;&gt;Claude Code 2.1.317–2.1.319: Background MCP Tasks and Deny-Instead-of-Prompt&lt;/h2&gt;
&lt;p&gt;Claude Code shipped three releases this week. Version 2.1.319 (November 6) moves an MCP tool call to the background when the server runs it as an MCP task and reports the result when it finishes. Opt out with &lt;code&gt;CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND&lt;/code&gt;, or disable MCP tasks entirely with &lt;code&gt;CLAUDE_CODE_DISABLE_MCP_TASKS=1&lt;/code&gt;. A new &lt;code&gt;permissions.denyInsteadOfPrompting&lt;/code&gt; setting denies any tool call that would otherwise wait on a permission prompt and lets the turn continue. That is useful for unattended terminal sessions, subagents and teammates that should never block.&lt;/p&gt;
&lt;p&gt;The same release adds &lt;code&gt;/claude-api latency-optimize&lt;/code&gt; and a &lt;code&gt;CLAUDE_CODE_DISABLE_FEATURE_FLAGS&lt;/code&gt; variable that stops feature-flag fetching without disabling telemetry. Auto mode now asks about or denies calls it cannot check with the server when telemetry or flags are off, which includes setups behind LLM gateways. &lt;code&gt;CLAUDE_CODE_AUTO_MODE_SERVER=0&lt;/code&gt; restores the built-in classifier. A self-paced &lt;code&gt;/loop&lt;/code&gt; now ends when a turn finishes without rescheduling.&lt;/p&gt;
&lt;p&gt;Version 2.1.318 (November 4) adds &lt;code&gt;loginExpiresAt&lt;/code&gt; to &lt;code&gt;claude auth status&lt;/code&gt; and &lt;code&gt;credential_expires_at&lt;/code&gt; to &lt;code&gt;claude -p&lt;/code&gt; JSON output, and stops auto-installing the IDE extension when launched from an IDE terminal. Version 2.1.317 (November 3) prompts before a web or shell tool call sends the account email address to the network, adds &lt;code&gt;onFailure: &amp;quot;block&amp;quot;&lt;/code&gt; for Agent SDK hook callbacks, and adds a cost report endpoint to the Claude apps gateway admin API.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Read more&lt;/a&gt; — Claude Code Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;codex-cli-0.167.0-adds-an-agent-command-center&quot; tabindex=&quot;-1&quot;&gt;Codex CLI 0.167.0 Adds an Agent Command Center&lt;/h2&gt;
&lt;p&gt;OpenAI released Codex CLI 0.167.0 on November 6. The headline feature is &lt;code&gt;codex agents --no-daemon&lt;/code&gt;, which opens an agent command center listing running and finished agent tasks, with unread results marked. Users can also switch between connected backends from the terminal session picker without losing drafts, and preview local image attachments while composing.&lt;/p&gt;
&lt;p&gt;Sign-in alternatives now sit under &amp;quot;Sign in another way&amp;quot;. A &lt;code&gt;gcp_setup_wizard&lt;/code&gt; flag configures Google Cloud Vertex AI as a provider with Application Default Credentials or a credential file, following the Vertex provider added in 0.165. There is also experimental opt-in support for macOS and Windows system proxy settings. Fixes include &lt;code&gt;Ctrl+L&lt;/code&gt; clearing the display while keeping history reachable via &lt;code&gt;Ctrl+T&lt;/code&gt;, and network approval prompts resolving when their underlying requests end.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/openai/codex/releases/tag/rust-v0.167.0&quot;&gt;Read more&lt;/a&gt; — OpenAI Codex (GitHub)&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;microsoft-365-copilot%3A-cowork-drives-your-local-edge%2C-admins-get-an-agents-usage-report&quot; tabindex=&quot;-1&quot;&gt;Microsoft 365 Copilot: Cowork Drives Your Local Edge, Admins Get an Agents Usage Report&lt;/h2&gt;
&lt;p&gt;Microsoft&#39;s November 4 release notes, covering October 28 to November 3, include several agent-related changes. Copilot Cowork can now use the user&#39;s local Microsoft Edge to complete web steps in a separate tab, using existing signed-in sessions while credentials stay on the device. That pattern resembles browser-use agents in other assistants, and it is worth reviewing against your organization&#39;s browser policies.&lt;/p&gt;
&lt;p&gt;For administrators, the Microsoft 365 admin center adds an agents usage report for Copilot and Copilot Chat, segmented by licensed vs. unlicensed users and by publisher type. Admins can also designate up to 100 SharePoint sites as authoritative content for Copilot Search. Other items include a SharePoint list agent that generates Microsoft Lists schemas from natural-language prompts, a HubSpot connector for Researcher, PowerPoint creating decks that reuse an attached presentation&#39;s theme, and Copilot-powered scheduled prompts in the Teams Workflows app (off by default).&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;Read more&lt;/a&gt; — Microsoft Learn&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;hermes-agent-v0.23.0-%22the-hearth-release%22-runs-the-desktop-workspace-in-a-browser&quot; tabindex=&quot;-1&quot;&gt;Hermes Agent v0.23.0 &amp;quot;The Hearth Release&amp;quot; Runs the Desktop Workspace in a Browser&lt;/h2&gt;
&lt;p&gt;NousResearch released Hermes Agent v0.23.0 on November 5. Its main feature is a Webapp that serves the full Desktop workspace in a browser tab from the user&#39;s own gateway. It is off by default and enabled via &lt;code&gt;webapp.enabled&lt;/code&gt; or Web Access in Settings. Architecturally, one gateway now owns every session, so the CLI, TUI, Desktop, ACP editors, chat bots and cron jobs all attach to it as clients.&lt;/p&gt;
&lt;p&gt;Other additions include &lt;code&gt;/rewind&lt;/code&gt; (press Esc twice), inline images in the terminal, pluggable delegation via &lt;code&gt;delegation.backend&lt;/code&gt; with an official &lt;code&gt;advanced-delegation&lt;/code&gt; plugin, &lt;code&gt;hermes plugins update&lt;/code&gt; with opt-in auto-update, and price tags and spend limits for paid MCP tools. GLM-5.3 is now the default model, and the plugin catalog grew from about 600 to over 1,050 entries.&lt;/p&gt;
&lt;p&gt;There are breaking changes to check before upgrading. The old built-in browser tools are gone in favor of &lt;code&gt;browser_exec&lt;/code&gt;, &lt;code&gt;hermes mcp serve&lt;/code&gt; is renamed to &lt;code&gt;hermes mcp-serve&lt;/code&gt;, Desktop requires macOS 13+ or 64-bit Windows, and built-in skills shrank from 58 to 44 as the Word, PowerPoint and Excel skills merged into a single &lt;code&gt;microsoft-office&lt;/code&gt; skill.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.0&quot;&gt;Read more&lt;/a&gt; — NousResearch (GitHub)&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;google-deepmind-releases-cheminformatics-skills-for-antigravity&quot; tabindex=&quot;-1&quot;&gt;Google DeepMind Releases Cheminformatics Skills for Antigravity&lt;/h2&gt;
&lt;p&gt;Google DeepMind published Cheminformatics Skills for Google Antigravity, an agent skill pack grounded in the RDKit toolkit and connected to the ChEMBL and PubChem databases. The walkthrough has the agent rigidify the kinase inhibitor crizotinib into lorlatinib-style macrocycles. Along the way, the agent retrieved thousands of compounds, curated 568 into a dashboard, ran matched molecular pair analysis, and trained a potency model using scaffold splits to avoid leakage.&lt;/p&gt;
&lt;p&gt;The skills are available now through Antigravity&#39;s Science plugin, and Google is offering 10,000 scientists the highest usage tier free for six months. The post notes that generated designs still need experimental validation. For developers, it is a worked example of domain skills: deterministic, well-tested libraries wrapped as agent tools so the model orchestrates them instead of making up chemistry.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/&quot;&gt;Read more&lt;/a&gt; — Google DeepMind&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-03-github-copilot-is-now-available-in-jetbrains-air&quot;&gt;GitHub Copilot is now available in JetBrains Air&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-04-github-copilot-is-now-available-in-xcodes-coding-assistant&quot;&gt;GitHub Copilot is now available in Xcode&#39;s coding assistant&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-06-github-copilot-weekly-releases-november-2&quot;&gt;GitHub Copilot weekly releases — November 2&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-11-06-github-copilot-in-slack-and-teams-is-generally-available&quot;&gt;@GitHub in Slack and Teams is generally available&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Claude Code Changelog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/openai/codex/releases/tag/rust-v0.167.0&quot;&gt;Codex CLI 0.167.0&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;Microsoft 365 Copilot release notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.0&quot;&gt;Hermes Agent v0.23.0&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/&quot;&gt;Agentic cheminformatics with Google Antigravity&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>GitHub Copilot now runs inside Xcode&#39;s built-in coding assistant and as an ACP agent in JetBrains Air, and the @GitHub bot for Slack and Teams went GA.</summary>
    
  </entry>
  
  <entry>
    <title>Cloud &amp; Infrastructure News: Step Functions as Agent Tools, S3 Tables Migrates Hive to Iceberg, Gemini Omni 1.2 Flash GA, Gemini Code Assist Entitlement Ends, 2026-11-08</title>
    <link href="https://javarubberduck.com/cloud/news-2026-11-08-cloud/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/cloud/news-2026-11-08-cloud/</id>
    <content type="html">&lt;h2 id=&quot;aws-weekly-roundup%3A-step-functions-as-agent-tools%2C-gpt-6.1-terra-on-bedrock%2C-lambda-soci&quot; tabindex=&quot;-1&quot;&gt;AWS Weekly Roundup: Step Functions as Agent Tools, GPT-6.1 Terra on Bedrock, Lambda SOCI&lt;/h2&gt;
&lt;p&gt;The November 2 AWS Weekly Roundup leads with OpenAI&#39;s GPT-6.1 Terra being generally available on Amazon Bedrock. AWS says input and output token prices are at least 50% lower than GPT-6.1 Sol and GPT-5.6 Terra, with up to 1 million input tokens and text plus image input.&lt;/p&gt;
&lt;p&gt;The more interesting item architecturally is that AWS Step Functions workflows, both Standard and Express, can now be exposed as AI agent tools through Amazon Bedrock AgentCore Gateway over MCP. It is available in every Region where AgentCore Gateway runs, at standard pricing. Instead of letting an agent string together many fine-grained API calls, teams can wrap a vetted, auditable multi-step workflow as a single tool. Retries, error handling and IAM boundaries then stay in the state machine rather than in the model&#39;s reasoning.&lt;/p&gt;
&lt;p&gt;Other items in the roundup:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;AWS Lambda&lt;/strong&gt; supports SOCI lazy loading for Lambda Managed Instances, with AWS citing up to 70% faster container startup. SOCI indexes are pushed to Amazon ECR.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Amazon ECS&lt;/strong&gt; adds GPU sharing with NVIDIA MPS on ECS Managed Instances and ECS on EC2.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Amazon EKS Auto Mode&lt;/strong&gt; supports the Kubernetes Gateway API at no extra charge beyond standard load balancing.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;AWS KMS&lt;/strong&gt; adds ML-KEM post-quantum key encapsulation.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-gpt-6-1-terra-on-bedrock-step-functions-as-agent-tools-and-more-november-2-2026/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;amazon-s3-tables-offers-managed-migration-from-apache-hive-to-apache-iceberg&quot; tabindex=&quot;-1&quot;&gt;Amazon S3 Tables Offers Managed Migration from Apache Hive to Apache Iceberg&lt;/h2&gt;
&lt;p&gt;Amazon S3 Tables can now migrate Parquet tables registered in the AWS Glue Data Catalog from Hive format to Apache Iceberg as a managed operation. There are no pipelines to build or compute to provision, and production workloads keep running during the migration. The process validates the source, produces a compatibility report, then copies data. You can run incremental syncs as often as needed, and the final cutover takes only a few minutes of write downtime.&lt;/p&gt;
&lt;p&gt;The destination table is read-only while migration is in progress and becomes read-write on completion. You pay only for the data copied, with no charge for the conversion itself. The CLI calls are &lt;code&gt;aws s3tables start-table-migration&lt;/code&gt; and &lt;code&gt;aws s3tables sync-table-migration&lt;/code&gt;, and Amazon Athena and SageMaker Unified Studio integrate directly. At launch the feature supports same-account, same-Region migrations in eleven Regions across the US, Asia Pacific and Europe.&lt;/p&gt;
&lt;p&gt;For data platform teams with large Hive estates, this removes much of the custom Spark work usually needed to move to Iceberg and its ACID transactions, schema evolution and time travel.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/amazon-s3-tables-now-offer-managed-migration-of-apache-hive-tables-to-apache-iceberg/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gemini-omni-1.2-flash-is-generally-available-for-video-generation-and-editing&quot; tabindex=&quot;-1&quot;&gt;Gemini Omni 1.2 Flash Is Generally Available for Video Generation and Editing&lt;/h2&gt;
&lt;p&gt;Google made Gemini Omni 1.2 Flash generally available on Google Cloud on November 6, with API access under the model ID &lt;code&gt;gemini-omni-1.2-flash&lt;/code&gt;. Pricing is $0.07 per second of 720p video, which Google says is 30% less than Omni 1.1 Flash. Maximum clip length goes from 10 to 15 seconds. Drafts can be generated at 360p or 720p and then upscaled to 1080p or 4K, which keeps iteration cheap.&lt;/p&gt;
&lt;p&gt;The model supports Provisioned Throughput for predictable capacity, and all outputs carry a SynthID watermark. Audio references are gated behind allowlisting. In the &lt;code&gt;global&lt;/code&gt; region of Gemini Enterprise, Omni 1.2 Flash replaces Veo 3.1 for video generation, so teams calling Veo through Gemini Enterprise should check their outputs and quotas.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://cloud.google.com/blog/products/ai-machine-learning/gemini-omni-1-2-flash-generally-available&quot;&gt;Read more&lt;/a&gt; — Google Cloud Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gemini-enterprise-ends-bundled-gemini-code-assist-entitlements&quot; tabindex=&quot;-1&quot;&gt;Gemini Enterprise Ends Bundled Gemini Code Assist Entitlements&lt;/h2&gt;
&lt;p&gt;Google Cloud&#39;s November 6 release notes announce that bundled Gemini Code Assist entitlements for existing Gemini Enterprise Standard and Plus subscriptions end on March 31, 2027. Until then, support is limited to the Gemini 3.x model family. New Gemini Enterprise subscriptions have not included a Code Assist entitlement since October 9, 2026.&lt;/p&gt;
&lt;p&gt;Google points developers to Google Antigravity, which is included in Gemini Enterprise editions, as the replacement for agentic coding. Organizations that rolled out Code Assist IDE extensions through Gemini Enterprise should plan their migration this quarter. That includes re-evaluating IDE plugins, admin policies and any code-customization indexes built on Code Assist.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud Release Notes&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-gpt-6-1-terra-on-bedrock-step-functions-as-agent-tools-and-more-november-2-2026/&quot;&gt;AWS Weekly Roundup: GPT-6.1 Terra on Bedrock, Step Functions as agent tools, and more (November 2, 2026)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/amazon-s3-tables-now-offer-managed-migration-of-apache-hive-tables-to-apache-iceberg/&quot;&gt;Amazon S3 Tables now offer managed migration of Apache Hive tables to Apache Iceberg&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://cloud.google.com/blog/products/ai-machine-learning/gemini-omni-1-2-flash-generally-available&quot;&gt;Gemini Omni 1.2 Flash generally available&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/release-notes&quot;&gt;Google Cloud release notes&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>AWS Step Functions workflows can now be exposed as MCP tools for AI agents through Bedrock AgentCore Gateway, and GPT-6.1 Terra is GA on Amazon Bedrock.</summary>
    
  </entry>
  
  <entry>
    <title>AI Dev Patterns: Hugging Face&#39;s Model Integration Contracts, CWE-bench v2, GitLab 19.5 Governs External Agents, 2026-11-08</title>
    <link href="https://javarubberduck.com/ai/news-2026-11-08-ai-patterns/"/>
    <updated>2026-11-08T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-11-08-ai-patterns/</id>
    <content type="html">&lt;h2 id=&quot;hugging-face-proposes-model-integration-contracts-for-transformers&quot; tabindex=&quot;-1&quot;&gt;Hugging Face Proposes Model Integration Contracts for Transformers&lt;/h2&gt;
&lt;p&gt;Hugging Face published a postmortem-driven proposal for testing model integrations at scale. On July 30, 2026, Transformers enabled PyTorch SDPA attention for T5. A side effect was that UMT5&#39;s decoder could attend to future tokens during training. The bug sat in main for almost two months before a fix on September 28. With 750+ supported model types, 88,000+ CPU tests on every main commit and 200,000+ GPU tests daily, full integration testing on every PR is not feasible.&lt;/p&gt;
&lt;p&gt;The proposed &lt;code&gt;transformers-mic&lt;/code&gt; standard packages a Model Integration Contract with three parts: a shrunk fixture model, a runnable usage path, and human-reviewed baselines. The UMT5 fixture is 66 MB against a 1,227 MB checkpoint. The shrink recipe turned Llama 4 Scout (108.6B parameters) into a 44.7M-parameter fixture. Replayed against history, the UMT5 contract flagged the bug on every affected commit, with label log-probs off by 1.22. One lesson: the fixture has to use the default SDPA path, because eager attention missed the bug entirely.&lt;/p&gt;
&lt;p&gt;The data on fixture design is useful beyond Transformers. On Whisper, random-weight fixtures caught 18 of 24 injected faults, while slightly overfit fixtures caught all 24. Since November 1, a MIC job in Transformers CI reports failures without blocking PRs. The CLI is &lt;code&gt;tmic&lt;/code&gt;, installed with &lt;code&gt;pip install &amp;quot;transformers-mic[runtime]&amp;quot;&lt;/code&gt;. Any team maintaining model-serving or fine-tuning code can apply the same pattern of small, deterministic, behavior-level contracts.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/transformers-mic&quot;&gt;Read more&lt;/a&gt; — Hugging Face&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;cwe-bench-v2-measures-whether-models-can-fix-real-vulnerability-classes&quot; tabindex=&quot;-1&quot;&gt;CWE-bench v2 Measures Whether Models Can Fix Real Vulnerability Classes&lt;/h2&gt;
&lt;p&gt;Collinear AI released CWE-bench v2 on Hugging Face. It is a security-repair benchmark with a public set of 100 tasks spanning 91 CWE categories, plus a separate held-out set to limit contamination. Tasks run through the Harbor evaluation harness (&lt;code&gt;harbor run ... -d cwe-bench@2.0 -k 4&lt;/code&gt;), and the dataset is gated.&lt;/p&gt;
&lt;p&gt;Of the 11 models evaluated, Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, followed closely by Claude Opus 5.5 at 67.3% pass@1 and 78.0% pass@4. On cost efficiency, DeepSeek V4.1 Flash reaches 57.0% pass@1 at about $0.04 per rollout. Even frontier models fail roughly a third of single-attempt vulnerability fixes, so AI-generated security patches still need the same review and testing as human ones.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/cwe-bench&quot;&gt;Read more&lt;/a&gt; — Hugging Face&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gitlab-19.5-makes-external-coding-agents-a-governed-part-of-the-platform&quot; tabindex=&quot;-1&quot;&gt;GitLab 19.5 Makes External Coding Agents a Governed Part of the Platform&lt;/h2&gt;
&lt;p&gt;InfoQ&#39;s analysis of GitLab 19.5 highlights a shift from &amp;quot;bring your own agent, unsupervised&amp;quot; to platform-governed agent use. The GitLab MCP server is now generally available, and admins can choose which MCP toolsets are exposed. Rate limits on &lt;a href=&quot;http://gitlab.com/&quot;&gt;GitLab.com&lt;/a&gt; are 60 requests per minute for Free and 600 for Premium and Ultimate, and reconnects and retries count toward them.&lt;/p&gt;
&lt;p&gt;Externally Connected Agents such as Claude Code, Cursor and Codex connect through the GitLab CLI. Qualifying sessions are recorded with user, agent, project and timing in an AI Governance Dashboard, and audit events can stream to a SIEM. The GitLab Duo CLI gained an auto mode that runs tools without per-command approval. It is disabled by default, configurable at instance, group and project levels, and tools marked Always Deny stay blocked. GitLab Secrets Manager, built on OpenBao with job-scoped secrets, also reached GA.&lt;/p&gt;
&lt;p&gt;The pattern of an MCP surface with toolset allow-lists, per-agent audit trails and tiered auto-approval is converging across DevOps platforms. It is a reasonable template for teams building internal agent access policies.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/gitlab-19-5-agentic-platform/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;where-should-an-agent&#39;s-code-live%3F-docker-compares-mounts%2C-worktrees-and-clone-mode&quot; tabindex=&quot;-1&quot;&gt;Where Should an Agent&#39;s Code Live? Docker Compares Mounts, Worktrees and Clone Mode&lt;/h2&gt;
&lt;p&gt;A Docker post compares three ways to give a sandboxed coding agent access to a repository with the &lt;code&gt;sbx&lt;/code&gt; CLI. Mounting the host repo (the default for &lt;code&gt;sbx run&lt;/code&gt; in a project directory) shows host edits immediately and lets host dev servers hot-reload, but dependencies are shared. Native binaries such as esbuild can then break because the host is macOS while the sandbox is Linux.&lt;/p&gt;
&lt;p&gt;Git worktrees give each task its own working directory and sandbox, which suits parallel agents. However, a linked worktree&#39;s &lt;code&gt;.git&lt;/code&gt; file points outside the mount, so &lt;code&gt;git status&lt;/code&gt;, branching and commits must run on the host. Clone mode mounts the host source read-only and gives the agent an independent clone. Work comes back through a Git remote that &lt;code&gt;sbx&lt;/code&gt; adds on the host, or by pushing to origin. That gives the strongest separation and a natural review gate. A fully mountless sandbox clones inside the VM, and anything left only there is deleted with the sandbox.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.docker.com/blog/docker-sandbox-code-checkout-workflows/&quot;&gt;Read more&lt;/a&gt; — Docker Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h3 id=&quot;safe-%26-secure-ai-agent-practices&quot; tabindex=&quot;-1&quot;&gt;Safe &amp;amp; Secure AI Agent Practices&lt;/h3&gt;
&lt;h2 id=&quot;docker-sandboxes-vs.-containers%3A-why-unattended-agents-need-a-separate-kernel&quot; tabindex=&quot;-1&quot;&gt;Docker Sandboxes vs. Containers: Why Unattended Agents Need a Separate Kernel&lt;/h2&gt;
&lt;p&gt;Docker published guidance on choosing isolation for AI agents. The core argument is that a container shares the host kernel, so a container escape or a mounted Docker socket gives a misbehaving agent a path to the host. Docker Sandboxes are microVMs with their own Linux kernel, filesystem, network stack and private Docker Engine. The agent can have sudo inside the VM and run Docker Compose or Testcontainers without ever touching the host socket.&lt;/p&gt;
&lt;p&gt;Sandboxes route traffic through a host-side proxy with a deny-by-default network policy and inject credentials on the host side, so secrets never sit inside the agent&#39;s environment. The workspace is mounted read-write by default, with clone mode as the stricter alternative. A sandboxed Claude Code session starts with &lt;code&gt;sbx run claude&lt;/code&gt;. Requirements are Apple silicon on macOS 14+, Windows Hypervisor Platform, or KVM on Ubuntu.&lt;/p&gt;
&lt;p&gt;Docker&#39;s recommendation is containers for interactive, supervised work and sandboxes for anything unattended. That matches the broader shift toward VM-grade isolation for autonomous agents.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.docker.com/blog/docker-sandboxes-vs-containers/&quot;&gt;Read more&lt;/a&gt; — Docker Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;edera&#39;s-object-capability-bounded-authority-keeps-credentials-out-of-agent-workloads&quot; tabindex=&quot;-1&quot;&gt;Edera&#39;s Object Capability Bounded Authority Keeps Credentials Out of Agent Workloads&lt;/h2&gt;
&lt;p&gt;Edera launched Object Capability Bounded Authority, a security model that isolates AI agent workloads and MCP servers from the credentials they act with, enforced at the hypervisor level. The idea is to separate the authority a task needs (&amp;quot;modify this one GitHub repository&amp;quot;) from the credential needed to exercise it (a personal access token). A compromised agent then has no static secret it can exfiltrate or reuse to move laterally.&lt;/p&gt;
&lt;p&gt;The Agent Runtime components are capcurl (credential-free API access through an isolated broker), capsudo (scoped command execution without standing root), and SPIFFE/SPIRE workload identity with mutual TLS instead of long-lived secrets. Object capabilities can also restrict an MCP server to specific external resources and commands. Scoped command execution and workload identity are released now. Credential-brokered API access and deeper orchestrator integrations will follow under Project Lunchbox.&lt;/p&gt;
&lt;p&gt;The claims are vendor-reported, but the pattern itself is sound and increasingly common. Keep secrets out of environment variables and config files the agent can read, and broker access per action instead.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/ai-credentials-isolation/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;github-actions-lockfiles-reach-ga%2C-but-do-not-make-dependencies-safe&quot; tabindex=&quot;-1&quot;&gt;GitHub Actions Lockfiles Reach GA, but Do Not Make Dependencies Safe&lt;/h2&gt;
&lt;p&gt;GitHub Actions workflow dependency lockfiles became generally available on October 28, two weeks after their public preview. The runner executes exactly the commits recorded in &lt;code&gt;.github/workflows/actions.lock&lt;/code&gt;, even if a referenced tag or branch is later moved. That closes the tag-repointing attack used in several recent supply-chain compromises of popular actions. Lockfiles are available on Enterprise Cloud (including data residency), Team and Free plans.&lt;/p&gt;
&lt;p&gt;The &lt;code&gt;gh actions-lock&lt;/code&gt; CLI extension generates and maintains the lockfile, and its verification mode can serve as a CI gate. Dependabot updates the lockfile entry in the same PR as the workflow reference change, and admins can enforce a &amp;quot;Require lockfile&amp;quot; policy at enterprise, organization or repository level. A separate GitHub changelog says &lt;code&gt;pull_request_target&lt;/code&gt; will be blocked by default from December 2.&lt;/p&gt;
&lt;p&gt;InfoQ points out the limit: a lockfile guarantees which code runs, not that it is safe. Teams that auto-merge Dependabot PRs will lock a malicious update in just as faithfully, so human review of action bumps still matters. That increasingly applies to agent-authored PRs as well.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/github-actions-lockfiles-ga/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/transformers-mic&quot;&gt;Transformers Model Integration Contracts&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/cwe-bench&quot;&gt;CWE-bench v2: More coverage, quality, and transparency&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/gitlab-19-5-agentic-platform/&quot;&gt;GitLab 19.5 Makes Agentic Development a Governed Part of the DevOps Platform&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.docker.com/blog/docker-sandbox-code-checkout-workflows/&quot;&gt;Where should your code live in a Docker Sandbox?&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.docker.com/blog/docker-sandboxes-vs-containers/&quot;&gt;Docker Sandboxes vs. Containers: Choosing the Right Isolation for AI Agents&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/ai-credentials-isolation/&quot;&gt;Edera&#39;s Object Capability Bounded Authority Promises to Isolate Credentials from AI Agent Workloads&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/11/github-actions-lockfiles-ga/&quot;&gt;GitHub Actions Lockfiles Reach GA to Secure Workflow Dependencies&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Hugging Face proposes Model Integration Contracts, shrunk fixtures with reviewed baselines, after an attention bug sat in Transformers main for two months undetected.</summary>
    
  </entry>
  
  <entry>
    <title>Spring Ecosystem Update: Boot 4.2 and Framework 7.1 Hit RC1, Spring AI 2.1 Adds Bring-Your-Own Embeddings and Reasoning Replay, 2026-11-02</title>
    <link href="https://javarubberduck.com/java/news-2026-11-02-spring/"/>
    <updated>2026-11-02T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-11-02-spring/</id>
    <content type="html">&lt;h2 id=&quot;spring-boot-4.2%2C-framework-7.1-and-the-whole-portfolio-reach-rc1&quot; tabindex=&quot;-1&quot;&gt;Spring Boot 4.2, Framework 7.1 and the Whole Portfolio Reach RC1&lt;/h2&gt;
&lt;p&gt;The week of October 19 brought first release candidates for almost every Spring project, which puts the November release train in its final stretch. Spring Boot 4.2.0-RC1 migrates the default buildpack builder to Ubuntu Resolute (replacing the Noble Java Tiny builder) and speeds up executable uber-JARs by looking up classes and resources in nested JARs directly. Spring Framework 7.1.0-RC1 improves nullability in &lt;code&gt;DataAccessUtils&lt;/code&gt; and &lt;code&gt;LinkedCaseInsensitiveMap&lt;/code&gt;, adds &lt;code&gt;onRetrySuccess()&lt;/code&gt; and &lt;code&gt;onRetryFailure()&lt;/code&gt; callbacks to &lt;code&gt;RetryListener&lt;/code&gt; that receive a &lt;code&gt;RetryState&lt;/code&gt;, and removes previously deprecated APIs.&lt;/p&gt;
&lt;p&gt;Several RCs carry changes worth checking before upgrading:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Spring Cloud 2026.0.0-RC1 (&amp;quot;Paddington&amp;quot;)&lt;/strong&gt; is based on Boot 4.2.0-RC1, and Spring Cloud Bus is no longer a separate project; it has been merged into Spring Cloud Stream.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Security 7.2.0-RC1&lt;/strong&gt; normalizes HTTP method casing in &lt;code&gt;StrictServerWebExchangeFirewall&lt;/code&gt;, and the opaque-token introspectors now send &lt;code&gt;application/x-www-form-urlencoded&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Modulith 2.2.0-RC1&lt;/strong&gt; adds &lt;code&gt;ABANDONED&lt;/code&gt; and &lt;code&gt;RESUBMITTED&lt;/code&gt; publication statuses and now completes event publications via &lt;code&gt;TransactionSynchronization&lt;/code&gt; inside the listener&#39;s transaction.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Batch 6.1.0-RC1&lt;/strong&gt; introduces a &lt;code&gt;@JobParameter&lt;/code&gt; annotation for injecting job parameters without expression-language strings, plus a type-safe &lt;code&gt;PageableItemReader&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Spring Integration 7.2.0-RC1&lt;/strong&gt; deprecates the default constructor of &lt;code&gt;SerializingHttpMessageConverter&lt;/code&gt; in favor of one that takes an allow-list of patterns.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Spring Data 2026.1.0-RC1, Spring for GraphQL 2.1.0-RC1, Spring Session 4.2.0-RC1, Spring AMQP and Spring for Apache Kafka 4.2.0-RC1, Spring LDAP 4.2.0-RC1 (with breaking changes) and Spring AI 2.1.0-RC1 shipped the same week. RC1 is the right moment to run your test suite against the new versions, since API changes after this point should be limited to fixes.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/spring-news-roundup-oct19-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;spring-ai-2.1-lets-you-bring-your-own-embeddings&quot; tabindex=&quot;-1&quot;&gt;Spring AI 2.1 Lets You Bring Your Own Embeddings&lt;/h2&gt;
&lt;p&gt;Until now, &lt;code&gt;VectorStore.add()&lt;/code&gt; always called the store&#39;s configured embedding model. Spring AI 2.1 adds &lt;code&gt;VectorStore.upsert(List&amp;lt;EmbeddedDocument&amp;gt;)&lt;/code&gt; for writing documents together with precomputed &lt;code&gt;float[]&lt;/code&gt; vectors, and &lt;code&gt;similaritySearch(float[], SearchRequest)&lt;/code&gt; for searching with a query vector directly. &lt;code&gt;topK&lt;/code&gt;, similarity thresholds and metadata filters still apply. &lt;code&gt;EmbeddedDocument&lt;/code&gt; rejects empty vectors and non-finite values at construction time.&lt;/p&gt;
&lt;p&gt;The Spring team&#39;s example is multimodal product search: embed product photos with &lt;code&gt;VertexAiMultimodalEmbeddingModel&lt;/code&gt;, store them with a stable ID derived from the SKU and a URL reference to the image in metadata, then embed a text query such as &amp;quot;green velvet armchair&amp;quot; with the same model and search by vector. The same mechanism lets teams reuse vectors produced by batch embedding jobs or migrations instead of paying to re-embed text.&lt;/p&gt;
&lt;p&gt;Stores validate vector dimensions before writing when the index dimension is known, and Redis refuses to write if it cannot determine the expected size. Support in 2.1.0-RC1 covers PGvector, Redis, Elasticsearch, Qdrant and OpenSearch; other stores throw &lt;code&gt;UnsupportedOperationException&lt;/code&gt; from default methods until implemented. An &lt;code&gt;upsert&lt;/code&gt; call is not transactional, so a partial failure should be retried with the same IDs. The post recommends storing the embedding model name and version in metadata so you know which documents need regenerating after a model change.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/28/spring-ai-upsert-own-embeddings&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;spring-ai-2.1-message-parts-keep-reasoning-intact-across-turns&quot; tabindex=&quot;-1&quot;&gt;Spring AI 2.1 Message Parts Keep Reasoning Intact Across Turns&lt;/h2&gt;
&lt;p&gt;Spring AI 2.1 changes the message data model: every message now holds an ordered &lt;code&gt;List&amp;lt;MessagePart&amp;gt;&lt;/code&gt; instead of a text string plus separate lists for tool calls and media. &lt;code&gt;MessagePart&lt;/code&gt; is a sealed interface with &lt;code&gt;TextPart&lt;/code&gt;, &lt;code&gt;ReasoningPart&lt;/code&gt;, &lt;code&gt;ToolCallPart&lt;/code&gt;, &lt;code&gt;ToolResultPart&lt;/code&gt;, &lt;code&gt;MediaPart&lt;/code&gt; and &lt;code&gt;UnknownPart&lt;/code&gt; records, and provider-issued signatures or encrypted reasoning are carried as an &lt;code&gt;OpaquePayload&lt;/code&gt; that is only replayed to the provider that issued it. Existing accessors such as &lt;code&gt;getText()&lt;/code&gt; and &lt;code&gt;getToolCalls()&lt;/code&gt; become views over the parts, and a new &lt;code&gt;getReasoning()&lt;/code&gt; replaces provider-specific reasoning access.&lt;/p&gt;
&lt;p&gt;This matters for tool-calling loops with reasoning models. Anthropic rejects requests where thinking blocks were altered, Gemini 3+ function calls fail without their thought signatures, and misordered replays can cause HTTP 400 errors or silently drop earlier reasoning. Spring AI 2.1 now replays reasoning correctly for Anthropic, Google GenAI, OpenAI-compatible servers (vLLM, gpt-oss, Kimi), Ollama, Bedrock Converse, DeepSeek and Mistral. A new &lt;code&gt;OpenAiResponsesChatModel&lt;/code&gt; supports OpenAI&#39;s Responses API in stateless mode, mapping each response item to a part in order.&lt;/p&gt;
&lt;p&gt;Persistence moves to Spring AI Session, which stores conversations as an append-only event log of full message parts; &lt;code&gt;SessionMemoryAdvisor&lt;/code&gt; records every tool round, and compaction only cuts at turn boundaries. &lt;code&gt;ChatMemory&lt;/code&gt; and its repositories are deprecated and scheduled for removal in 3.0. Histories written before 2.1 contain no reasoning parts, and reasoning signed by one provider is never sent to another when you switch models.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/28/spring-ai-message-parts-reasoning-replay&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/spring-news-roundup-oct19-2026/&quot;&gt;Spring News Roundup: First Release Candidates of Boot, Framework, Data, Security, Modulith, AI&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/28/spring-ai-upsert-own-embeddings&quot;&gt;Using Your Own Embeddings in Spring AI 2.1&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/28/spring-ai-message-parts-reasoning-replay&quot;&gt;Spring AI 2.1 Message Parts: Reasoning That Survives the Round Trip&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Spring Boot 4.2.0-RC1, Spring Framework 7.1.0-RC1 and release candidates for nearly every Spring project landed ahead of the November GA train, with Spring Cloud Bus folded into Spring Cloud Stream.</summary>
    
  </entry>
  
  <entry>
    <title>Java News: October 2026 CPU Lands, eMudhra CA Keys Compromised, Value Classes Testing Call for JDK 28, 2026-11-02</title>
    <link href="https://javarubberduck.com/java/news-2026-11-02-java-jdk/"/>
    <updated>2026-11-02T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-11-02-java-jdk/</id>
    <content type="html">&lt;h2 id=&quot;october-2026-critical-patch-update-ships-across-all-supported-jdk-lines&quot; tabindex=&quot;-1&quot;&gt;October 2026 Critical Patch Update Ships Across All Supported JDK Lines&lt;/h2&gt;
&lt;p&gt;The quarterly Critical Patch Update for October 2026 is out, and InfoQ&#39;s roundup for the week of October 19 collects the vendor releases. Oracle shipped JDK 27.0.1, 25.0.5, 21.0.13, 17.0.21, 11.0.33 and 8u511. Azul released CPU builds of Zulu for JDK 27, 25, 21, 17, 11, 8, 7 and 6, covering 53 CVEs in total. BellSoft published Liberica JDK CPU builds (25.0.4.1.1, 21.0.12.1.1, 17.0.20.1.1, 11.0.32.1.1, 8u511, 7u521, 6u521) as well as Patch Set Update releases (27.0.1, 25.0.5, 21.0.13, 17.0.21, 11.0.33, 8u512), reporting 1,438 fixes and backports across all lines.&lt;/p&gt;
&lt;p&gt;The same week, the JDK 28 release schedule was proposed: Rampdown Phase One on December 3, 2026, Rampdown Phase Two on January 14, 2027, release candidates on February 4 and February 16, and General Availability on March 9, 2027. JDK 28 early-access Build 21 is also available. With Rampdown Phase One about a month away, the feature set (currently ten targeted JEPs, including Value Objects, Structured Concurrency finalization and AOT code compilation) is close to frozen.&lt;/p&gt;
&lt;p&gt;For teams on an LTS line, the practical step is the usual one: roll the CPU build into base images and CI toolchains. Teams on the non-LTS JDK 27 should note that 27.0.1 is the first update in that line.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct19-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;eight-emudhra-intermediate-ca-keys-compromised%2C-out-of-cycle-jdk-updates-follow&quot; tabindex=&quot;-1&quot;&gt;Eight eMudhra Intermediate CA Keys Compromised, Out-of-Cycle JDK Updates Follow&lt;/h2&gt;
&lt;p&gt;Mozilla disclosed that the private keys for eight eMudhra intermediate CA certificates had been compromised. Anyone holding those keys could issue certificates impersonating arbitrary websites or services. Browsers and platforms moved quickly: Firefox and Chrome blocked the intermediates on October 17, Microsoft platforms on October 20 and Apple products on October 22. Oracle published security alert CVE-2026-91403 for the Java runtime.&lt;/p&gt;
&lt;p&gt;Azul shipped an out-of-cycle Zulu update on October 21 specifically for this issue, and pointed out the detail that matters most to Java developers: the JDK does not check certificate revocation by default. A JVM with the affected intermediates still trusted will accept a fraudulent certificate until its trust store is updated, regardless of whether the CA has been revoked upstream.&lt;/p&gt;
&lt;p&gt;If your services make outbound TLS calls, confirm that every runtime (including those in container images that are not rebuilt often) has picked up the updated trust store or the vendor&#39;s out-of-cycle build. Applications that ship a custom &lt;code&gt;cacerts&lt;/code&gt; file need to remove the affected intermediates themselves.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct19-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;quality-outreach-asks-library-maintainers-to-test-jdk-28-value-classes&quot; tabindex=&quot;-1&quot;&gt;Quality Outreach Asks Library Maintainers to Test JDK 28 Value Classes&lt;/h2&gt;
&lt;p&gt;With JEP 401 (Value Objects, Preview) targeted to JDK 28, the OpenJDK Quality Outreach program sent a heads-up asking open source projects to test against JDK 28 in two configurations: with preview features disabled (behavior should be unchanged) and with preview features enabled. Projects that can do so are asked to run both in CI and report surprises to the valhalla-dev mailing list.&lt;/p&gt;
&lt;p&gt;The reason for the second configuration is that around 30 JDK classes become value classes when preview is enabled, even for code compiled without preview or on an older JDK. That includes the primitive wrappers, &lt;code&gt;Number&lt;/code&gt;, &lt;code&gt;Record&lt;/code&gt;, &lt;code&gt;Optional&lt;/code&gt; and its primitive variants, and most &lt;code&gt;java.time&lt;/code&gt; types. For these, &lt;code&gt;==&lt;/code&gt; compares field values rather than identity, synchronizing on an instance fails with a compile error or &lt;code&gt;IdentityException&lt;/code&gt;, &lt;code&gt;java.lang.ref&lt;/code&gt; APIs and &lt;code&gt;WeakHashMap&lt;/code&gt; reject them, and reflective &lt;code&gt;setAccessible&lt;/code&gt; on their fields throws.&lt;/p&gt;
&lt;p&gt;The note also flags a source-level change: compiling with preview enabled applies early-construction rules, so record constructors that touch &lt;code&gt;this&lt;/code&gt; before all fields are assigned (for example by calling an instance method) no longer compile. Library maintainers who lock on boxed values, key weak caches on &lt;code&gt;Integer&lt;/code&gt; or &lt;code&gt;LocalDate&lt;/code&gt;, or rely on identity of these types should start testing now, while there is still time to file issues before JDK 28 ships in March.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://inside.java/2026/10/27/quality-heads-up/&quot;&gt;Read more&lt;/a&gt; — Inside Java&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;wildfly-42%2C-apache-poi-6.0-and-micronaut-5.2.3-round-out-the-ecosystem-week&quot; tabindex=&quot;-1&quot;&gt;WildFly 42, Apache POI 6.0 and Micronaut 5.2.3 Round Out the Ecosystem Week&lt;/h2&gt;
&lt;p&gt;WildFly 42, released October 22, adds virtual-thread support to its Jakarta Concurrency implementation, so managed executors can run tasks on virtual threads. It also lets the Elytron &lt;code&gt;token-realm&lt;/code&gt; apply a principal transformer (such as &lt;code&gt;case-principal-transformer&lt;/code&gt;) to principals derived from JWTs, which simplifies mapping external identity-provider subjects to local users.&lt;/p&gt;
&lt;p&gt;Apache POI 6.0.0 reached GA with a JDK 17 baseline. &lt;code&gt;DataFormatter&lt;/code&gt; now uses cached values for formula cells by default, and deprecated methods have been removed, so projects upgrading from 5.x should check for compilation breaks and for spreadsheets whose formatted output depended on re-evaluating formulas.&lt;/p&gt;
&lt;p&gt;On the framework side, Micronaut 5.2.3 resolves 10 CVEs through Micronaut Core 5.2.17 and Micronaut Security 5.4.1, and the Open Liberty 26.0.0.11 beta adds workload identity and JWK rotation to its JWT Builder plus a configurable server quiesce timeout. JobRunr 9.0.2 makes the Quarkus &lt;code&gt;JobRunrStarter&lt;/code&gt; bean replaceable.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct19-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct19-2026/&quot;&gt;Java News Roundup: JDK 28 Schedule, Oracle CPUs, eMudhra Certs, Apache POI, WildFly, Azul, BellSoft&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://inside.java/2026/10/27/quality-heads-up/&quot;&gt;Quality Outreach Heads-up - JDK 28: Value Classes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://openjdk.org/projects/jdk/28/&quot;&gt;JDK 28 project page&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>The October 2026 Critical Patch Update shipped JDK 27.0.1, 25.0.5, 21.0.13, 17.0.21, 11.0.33 and 8u511 across Oracle, Azul and BellSoft, while a compromise of eight eMudhra intermediate CA keys forced out-of-cycle trust-store updates.</summary>
    
  </entry>
  
  <entry>
    <title>Developer Tools Digest: GitHub Universe Brings Copilot Fusion and Copilot in GitHub Enterprise, Mistral Large 4 Open Weights, Anthropic Files S-1, 2026-11-02</title>
    <link href="https://javarubberduck.com/ai/news-2026-11-02-dev-tools/"/>
    <updated>2026-11-02T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-11-02-dev-tools/</id>
    <content type="html">&lt;h2 id=&quot;github-copilot-fusion-routes-each-prompt-across-multiple-models&quot; tabindex=&quot;-1&quot;&gt;GitHub Copilot Fusion Routes Each Prompt Across Multiple Models&lt;/h2&gt;
&lt;p&gt;GitHub used Universe on October 28 to launch Copilot Fusion, a model-picker option that replaces choosing a single model with choosing an effort level (Standard, High or Max). For each prompt, Fusion picks a workflow: &lt;strong&gt;Single&lt;/strong&gt; (one model handles it), &lt;strong&gt;Cascade&lt;/strong&gt; (a cheaper model attempts first, a quality check accepts the result or escalates to a stronger model that starts fresh) or &lt;strong&gt;Critique&lt;/strong&gt; (one model drafts, a second reviews, the first revises). It combines the HydraFusion research preview with Copilot&#39;s existing auto model selection.&lt;/p&gt;
&lt;p&gt;GitHub reports that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at roughly half the average API cost, half the tokens and about 40% less time per task. Billing is per underlying model at standard rates, with a temporary 30% discount for paid plans. A companion change lets Fusion select local models.&lt;/p&gt;
&lt;p&gt;Fusion is available on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. Admins only expose models already allowed by enterprise policy, can disable specific models for Fusion alone, and can make Fusion the default through managed settings. Research-preview users of HydraFusion are migrated automatically.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;copilot-comes-bundled-with-github-enterprise%2C-and-the-copilot-sdk-moves-to-microsoft&quot; tabindex=&quot;-1&quot;&gt;Copilot Comes Bundled with GitHub Enterprise, and the Copilot SDK Moves to Microsoft&lt;/h2&gt;
&lt;p&gt;Also at Universe, GitHub changed how enterprises buy Copilot. Customers working with a GitHub or Microsoft account team can now get Copilot included for every licensed GitHub Enterprise user instead of buying separate seats. Users without an Enterprise license can get Copilot Standalone at $10 per user per month. Both have the same features and models as Copilot Enterprise but no per-seat monthly AI credits: usage is metered from the first request against the enterprise budget.&lt;/p&gt;
&lt;p&gt;When that budget runs out, code completion and next-edit suggestions keep working, while Copilot Chat, code review and cloud features stop until budget is available. Self-serve subscriptions are not yet eligible, with dates to come &amp;quot;in the coming weeks&amp;quot;; Copilot Business and Enterprise remain on sale. For platform teams, the thing to manage shifts from seat counts to the AI usage budget and its alerts.&lt;/p&gt;
&lt;p&gt;Separately, the GitHub Copilot SDK is being renamed the Microsoft Copilot SDK and will serve a shared &amp;quot;Microsoft Copilot harness&amp;quot; used across Microsoft and GitHub products. The repository will move to the Microsoft organization in phases. GitHub says no immediate changes are required, but package names and namespaces may change later, so teams building on the SDK should watch the repository for migration notes.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-enterprise-now-includes-access-to-github-copilot&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;agent-merge-and-shared-copilot-automations-push-copilot-further-into-the-cloud&quot; tabindex=&quot;-1&quot;&gt;Agent Merge and Shared Copilot Automations Push Copilot Further into the Cloud&lt;/h2&gt;
&lt;p&gt;Agent Merge, previously available in the Copilot app and VS Code, is now built into pull requests on &lt;a href=&quot;http://github.com/&quot;&gt;github.com&lt;/a&gt;. A &amp;quot;Watch and fix&amp;quot; button in the merge box tells Copilot to monitor the PR and fix selected blockers (failing Actions workflows, merge conflicts or both) by spinning up a cloud development environment. Combined with auto-merge, it keeps the PR mergeable until GitHub merges it. It can be enabled by default for all your PRs and requires a paid Copilot plan.&lt;/p&gt;
&lt;p&gt;Copilot Automations, which combine prompts, triggers and allowed tools into workflows that Copilot runs autonomously, gained a shared mode. Shared automations belong to a repository rather than a person: anyone with read access can see them, anyone with write access can edit them, credits bill to the repository owner, actions are attributed to Copilot, and they keep running if the author leaves. Examples include reproducing a bug when an issue is opened or capturing screenshots when a PR is marked ready for review.&lt;/p&gt;
&lt;p&gt;Shared automations are rolling out to Pro, Pro+, Max, Business and Enterprise plans. In organization-owned repositories an admin must enable the policy that allows the Copilot cloud agent to bill to the organization. Universe also brought GA of GitHub Agentic Workflows, covered in this week&#39;s AI Dev Patterns digest.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-build-shared-workflows-for-your-team-with-copilot-automations&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gpt-6.1-terra-arrives-in-copilot-and-codex%2C-codex-cli-adds-a-vertex-ai-provider&quot; tabindex=&quot;-1&quot;&gt;GPT-6.1 Terra Arrives in Copilot and Codex, Codex CLI Adds a Vertex AI Provider&lt;/h2&gt;
&lt;p&gt;OpenAI released GPT-6.1 Terra on October 26, positioned below GPT-6.1 Sol in cost and focused on image understanding and computer use across browsers, desktop and apps. In ChatGPT Work and the desktop app it is available on Plus, Pro, Business, Enterprise and Edu plans. GitHub made it generally available in Copilot the same day across VS Code, Visual Studio, Copilot CLI, the cloud agent, JetBrains, Xcode and Eclipse, reporting that it solved noticeably more terminal and C# tasks than GPT-5.6 Terra and works best on well-scoped tasks.&lt;/p&gt;
&lt;p&gt;Codex CLI 0.165.0 (October 27) adds a built-in Google Cloud Vertex AI provider, selected with &lt;code&gt;model_provider = &amp;quot;gcp&amp;quot;&lt;/code&gt;, plus a repository filter in the session picker, directory grouping in the agent command center and experimental attached thread forks in the app-server API. Changing reasoning effort mid-conversation is now on by default, and Memory v2 is the default for fresh profiles. Codex CLI 0.166.0 (October 29) adds keyboard-driven transcript selection with search, &lt;code&gt;codex queue --resume&lt;/code&gt;, &lt;code&gt;codex doctor --network-only&lt;/code&gt;, and clearer explanations when a safety block triggers, with suggested safer follow-up prompts.&lt;/p&gt;
&lt;p&gt;GPT-6.1 Terra is now in the bundled, Vertex AI and Amazon Bedrock catalogs, so the same model can be used through all three providers from Codex CLI.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;Read more&lt;/a&gt; — OpenAI&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;claude-code-2.1.311%E2%80%932.1.316%3A-write-boundaries%2C-subagent-effort-and-gateway-budgets&quot; tabindex=&quot;-1&quot;&gt;Claude Code 2.1.311–2.1.316: Write Boundaries, Subagent Effort and Gateway Budgets&lt;/h2&gt;
&lt;p&gt;Claude Code shipped several releases over the past week. Version 2.1.314 adds &lt;code&gt;permissions.blockWritesOutsideWorkingDirectories&lt;/code&gt;, which makes the Write, Edit and NotebookEdit tools refuse paths outside the working directories in every permission mode, a useful hard boundary for teams running agents in auto-accept modes. Version 2.1.316 adds &lt;code&gt;CLAUDE_CODE_SUBAGENT_EFFORT&lt;/code&gt; to set the default effort level for subagents and &lt;code&gt;CLAUDE_CODE_DISABLE_ULTRATHINK&lt;/code&gt; to turn off the ultrathink keyword.&lt;/p&gt;
&lt;p&gt;Most of the remaining changes target organizations running Claude through the Claude apps gateway:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;2.1.312:&lt;/strong&gt; &lt;code&gt;oidc.user_id_claim&lt;/code&gt; identifies users by a claim other than &lt;code&gt;sub&lt;/code&gt;, such as Entra ID&#39;s &lt;code&gt;oid&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2.1.314:&lt;/strong&gt; gateway upstreams accept a &lt;code&gt;groups&lt;/code&gt; list, so only members of listed IdP groups are routed to them.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2.1.316:&lt;/strong&gt; pooled group budgets let a group&#39;s members draw from a shared monthly budget via &lt;code&gt;pooled_budget&lt;/code&gt;, and &lt;code&gt;managed.policy_mode: layered&lt;/code&gt; applies every matching managed policy instead of only the first.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Smaller additions include syntax coloring for code blocks in VS Code, a Skills view in Claude Tag analytics showing how often each attached skill is used, effort-level selection for cloud routines, and self-hosted runner flags that keep idle sessions alive while background processes they started are still running.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Read more&lt;/a&gt; — Anthropic&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;mistral-large-4-released-as-a-1t-parameter-open-weight-moe&quot; tabindex=&quot;-1&quot;&gt;Mistral Large 4 Released as a 1T-Parameter Open-Weight MoE&lt;/h2&gt;
&lt;p&gt;Mistral released Mistral Large 4 (nicknamed &amp;quot;Le Chonk&amp;quot;) on Hugging Face on October 29. It is a multimodal Mixture-of-Experts model with about 1T total parameters and roughly 49B active per token (5 of 128 experts, one shared), a 1M-token context window, Multi-head Latent Attention with a sparse-attention indexer, and a 1.6B Pixtral vision encoder. It takes text and images and produces text.&lt;/p&gt;
&lt;p&gt;Mistral emphasizes security and agentic coding results: 93% on Cybench, 82% on vulnerability reproduce-and-patch tasks in the Artificial Analysis Cyber Index, 93.3% attack resistance on Lakera B3, 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. It ranked as the fifth-best open-weight model on the Artificial Analysis Intelligence Index at launch.&lt;/p&gt;
&lt;p&gt;Official FP8 and NVFP4 checkpoints are available, along with community GGUF quantizations. It runs with Transformers across 8 GPUs, with vLLM on Blackwell for NVFP4, or via llama.cpp, where a Q2_K quant fits on a single 512 GB Apple Silicon machine; TRL supports LoRA fine-tuning. The license is the Mistral Ecosystem Open-Weight License, an MIT derivative that restricts offering the model as a service, so check it before hosting the model for third parties.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/mistral-large4&quot;&gt;Read more&lt;/a&gt; — Hugging Face&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;microsoft-365-copilot-adds-admin-built-mcp-connectors-and-usage-billed-work-iq-apis&quot; tabindex=&quot;-1&quot;&gt;Microsoft 365 Copilot Adds Admin-Built MCP Connectors and Usage-Billed Work IQ APIs&lt;/h2&gt;
&lt;p&gt;Microsoft&#39;s October 27 release notes for Microsoft 365 Copilot include two developer-relevant changes. Admins can now build custom connectors based on the Model Context Protocol in the Microsoft 365 admin center and deploy them across Microsoft 365, so organizational data from internal systems can surface in Copilot without building a full Graph connector.&lt;/p&gt;
&lt;p&gt;The Work IQ APIs gained usage-based billing, letting developers invoke agents and models without pre-assigned Copilot licenses, and a new endpoint provides programmatic access to both first-party and tenant-defined declarative agents. That makes it practical to call Copilot agents from internal applications where only a subset of users is licensed.&lt;/p&gt;
&lt;p&gt;The same release adds a Learning Agent, web links as sources in Copilot Notebooks, new Copilot usage metrics in Viva Insights, and PowerPoint generation that reuses an attached deck&#39;s theme.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;Read more&lt;/a&gt; — Microsoft Learn&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;anthropic-files-public-s-1-for-a-nasdaq-ipo-under-the-ticker-ant&quot; tabindex=&quot;-1&quot;&gt;Anthropic Files Public S-1 for a Nasdaq IPO Under the Ticker ANT&lt;/h2&gt;
&lt;p&gt;Anthropic, PBC announced on October 26 that it has filed a registration statement on Form S-1 with the SEC for a proposed initial public offering of Class A common stock. It intends to list on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley named as contacts for the preliminary prospectus. The number of shares and price range have not been determined, and the registration statement is not yet effective.&lt;/p&gt;
&lt;p&gt;The public filing follows the confidential S-1 Anthropic submitted in June. For developers the immediate impact is limited, but a public prospectus will disclose far more detail about Anthropic&#39;s revenue mix, compute commitments and risk factors than has been available, which is relevant context for teams making long-term bets on the Claude platform.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.anthropic.com/news/s1-sec&quot;&gt;Read more&lt;/a&gt; — Anthropic&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion&quot;&gt;Route and orchestrate models with GitHub Copilot Fusion&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-enterprise-now-includes-access-to-github-copilot&quot;&gt;GitHub Enterprise now includes access to GitHub Copilot&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-copilot-sdk-is-becoming-microsoft-copilot-sdk&quot;&gt;GitHub Copilot SDK is becoming Microsoft Copilot SDK&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-agent-merge-watches-and-fixes-pull-requests-in-the-cloud&quot;&gt;Agent Merge watches and fixes pull requests in the cloud&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-build-shared-workflows-for-your-team-with-copilot-automations&quot;&gt;Build shared workflows for your team with Copilot Automations&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-26-gpt-6-1-terra-in-github-copilot&quot;&gt;GPT-6.1 Terra in GitHub Copilot&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;OpenAI Codex changelog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Claude Code changelog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/mistral-large4&quot;&gt;Welcome Le Chonk: Mistral Large 4 on Hugging Face&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes?tabs=all&quot;&gt;Microsoft 365 Copilot release notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.anthropic.com/news/s1-sec&quot;&gt;Anthropic files S-1 with the SEC for proposed initial public offering&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>At GitHub Universe, Copilot Fusion became a picker option that routes each prompt across models via single, cascade or critique workflows, scoring within two points of Claude Code with Opus 5.5 Max at about half the cost per task.</summary>
    
  </entry>
  
  <entry>
    <title>Cloud &amp; Infrastructure News: TorchNeuron for Trainium, Structured Outputs for Claude on Bedrock, Gemini Computer Use GA, Managed OpenTelemetry for GKE, 2026-11-02</title>
    <link href="https://javarubberduck.com/cloud/news-2026-11-02-cloud/"/>
    <updated>2026-11-02T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/cloud/news-2026-11-02-cloud/</id>
    <content type="html">&lt;h2 id=&quot;torchneuron-brings-native-pytorch-to-aws-trainium&quot; tabindex=&quot;-1&quot;&gt;TorchNeuron Brings Native PyTorch to AWS Trainium&lt;/h2&gt;
&lt;p&gt;AWS released TorchNeuron in public beta alongside Neuron 2.33. It is an open source PyTorch backend that runs models on Trainium by switching the device to Neuron, without rewriting model code. Eager execution and &lt;code&gt;torch.compile&lt;/code&gt; both keep working, and it integrates with the Neuron Kernel Interface for custom operators.&lt;/p&gt;
&lt;p&gt;Supported hardware is Trainium2 and Trainium3. Previously, moving a PyTorch workload to Trainium meant adapting it to the Neuron-specific tracing workflow; a native backend makes Trainium much closer to a drop-in target for teams already training on GPUs, which matters for anyone evaluating Trainium&#39;s pricing for fine-tuning or post-training jobs.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-amazon-polly-generative-engine-torchneuron-for-trainium-and-more-october-26-2026/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;sagemaker-ai-training-sessions-and-structured-outputs-for-claude-on-bedrock&quot; tabindex=&quot;-1&quot;&gt;SageMaker AI Training Sessions and Structured Outputs for Claude on Bedrock&lt;/h2&gt;
&lt;p&gt;Amazon SageMaker AI training sessions are now generally available. They let you own the post-training loop (custom loss functions and reward signals) while SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage. Billing is per token processed, and charges pause while your reward function scores a batch, which makes reinforcement-style post-training with slow external graders cheaper. Muse Glimmer, Qwen 3.8 and Ministral 3 were added as supported base models.&lt;/p&gt;
&lt;p&gt;Amazon Bedrock now supports structured outputs for all Anthropic Claude models, returning JSON that conforms to a supplied schema on both the &lt;code&gt;bedrock-runtime&lt;/code&gt; and &lt;code&gt;bedrock-mantle&lt;/code&gt; endpoints. That removes the custom parsing and retry logic teams typically wrap around extraction or tool-style calls.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-amazon-polly-generative-engine-torchneuron-for-trainium-and-more-october-26-2026/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;elasticache-serverless-for-valkey-adds-zero-downtime-online-migration%3B-polly-generative-voices-expand&quot; tabindex=&quot;-1&quot;&gt;ElastiCache Serverless for Valkey Adds Zero-Downtime Online Migration; Polly Generative Voices Expand&lt;/h2&gt;
&lt;p&gt;Amazon ElastiCache Serverless for Valkey can now migrate self-managed Valkey or Redis OSS running on EC2 with no client-facing errors. The workflow is prepare, start data sync, then cut over manually or automatically, without third-party replication tools. It is available in all commercial Regions, the China Regions and GovCloud (US), which removes one of the main blockers for teams wanting to move off self-managed Redis clusters.&lt;/p&gt;
&lt;p&gt;Amazon Polly&#39;s Generative engine gained 14 voices across 9 locales, bringing the catalog to 57 generative voices in 11 languages and 25 locales. Hindi and Finnish are generally available for the first time, and the upgrade is drop-in: voice names, the &lt;code&gt;SynthesizeSpeech&lt;/code&gt; call and pricing are unchanged.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-amazon-polly-generative-engine-torchneuron-for-trainium-and-more-october-26-2026/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gemini-enterprise-agent-platform%3A-computer-use-tool-and-inference-metadata-logging-go-ga&quot; tabindex=&quot;-1&quot;&gt;Gemini Enterprise Agent Platform: Computer Use Tool and Inference Metadata Logging Go GA&lt;/h2&gt;
&lt;p&gt;On October 30, Google made the computer use tool generally available on Gemini Enterprise Agent Platform. It ships with human-in-the-loop safety policies: developers can configure or override specific confirmation policies, for example requiring approval before financial transactions or data modification, and administrators can enforce organization-wide guardrails through the &lt;code&gt;configurableComputerUseToolSafetyPolicies&lt;/code&gt; organization policy constraint.&lt;/p&gt;
&lt;p&gt;Inference request metadata logging also reached GA. It writes one Cloud Logging entry per request with model, region, consumption option, token counts, Provisioned Throughput usage, latency and error code, but never prompts or responses. It covers Gemini models via &lt;code&gt;generateContent&lt;/code&gt; and &lt;code&gt;streamGenerateContent&lt;/code&gt; and Claude models via &lt;code&gt;rawPredict&lt;/code&gt; and &lt;code&gt;streamRawPredict&lt;/code&gt;. Requests can be tagged with headers prefixed &lt;code&gt;x-goog-agent-platform-observability-label-&lt;/code&gt; and queried with SQL in Observability Analytics, which gives per-team or per-feature cost attribution without logging sensitive content.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;managed-opentelemetry-for-gke-is-generally-available&quot; tabindex=&quot;-1&quot;&gt;Managed OpenTelemetry for GKE Is Generally Available&lt;/h2&gt;
&lt;p&gt;Managed OpenTelemetry for GKE reached GA on October 29 for clusters on 1.37.1-gke.1697000 or later. It replaces a self-managed OpenTelemetry Collector with a Google-managed one that accepts OTLP traces, metrics and logs over gRPC on port 4317 as well as HTTP on 4318, configured through the &lt;code&gt;otlp_grpc&lt;/code&gt; setting of the Instrumentation custom resource. Logs now flow through the Cloud Telemetry API alongside traces and metrics, with improved per-signal queueing and autoscaling for high-throughput clusters.&lt;/p&gt;
&lt;p&gt;Also GA: multimodal prompt and response collection for LangGraph ReAct and Agent Development Kit agents, the &lt;code&gt;--managed-otel-scope&lt;/code&gt; gcloud flag (gcloud 576.0.0+), and &lt;code&gt;managed_opentelemetry_config&lt;/code&gt; in the Terraform Google provider (v8.8.0+). Clusters on Preview versions need upgrading to pick up the GA collector.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/kubernetes-engine/docs/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;cloud-run-jobs-can-run-gpus-for-up-to-12-hours%3B-alloydb-scales-vector-indexes-to-10-billion-rows&quot; tabindex=&quot;-1&quot;&gt;Cloud Run Jobs Can Run GPUs for Up to 12 Hours; AlloyDB Scales Vector Indexes to 10 Billion Rows&lt;/h2&gt;
&lt;p&gt;Cloud Run jobs can now run with GPUs attached for up to 12 hours, in Preview as of October 30. That makes serverless jobs viable for longer batch inference, embedding generation and fine-tuning runs that previously needed GKE or Compute Engine.&lt;/p&gt;
&lt;p&gt;On the data side, AlloyDB for PostgreSQL made four-level ScaNN vector indexes generally available on October 28, supporting up to 10 billion rows, and added &lt;code&gt;ai.extract&lt;/code&gt; and &lt;code&gt;ai.classify&lt;/code&gt; functions for in-query text processing in Preview. BigQuery made Git repositories for SQL scripts and notebooks and the &lt;code&gt;AI.COUNT_TOKENS&lt;/code&gt; function GA, and Spanner now supports equality and ordering comparisons on JSON values with search-index acceleration.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/run/docs/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-weekly-roundup-amazon-polly-generative-engine-torchneuron-for-trainium-and-more-october-26-2026/&quot;&gt;AWS Weekly Roundup: Amazon Polly Generative engine, TorchNeuron for Trainium, and more (October 26, 2026)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes&quot;&gt;Gemini Enterprise Agent Platform release notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/kubernetes-engine/docs/release-notes&quot;&gt;GKE release notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/run/docs/release-notes&quot;&gt;Cloud Run release notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/release-notes&quot;&gt;Google Cloud release notes&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>AWS released TorchNeuron in public beta, an open source PyTorch backend that runs existing models on Trainium2 and Trainium3 by switching the device to Neuron, with eager mode and torch.compile intact.</summary>
    
  </entry>
  
  <entry>
    <title>AI Dev Patterns: GitHub Agentic Workflows Go GA, Anthropic Replaces Git with Scion, Validator-Gated Security Agents, 2026-11-02</title>
    <link href="https://javarubberduck.com/ai/news-2026-11-02-ai-patterns/"/>
    <updated>2026-11-02T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-11-02-ai-patterns/</id>
    <content type="html">&lt;h2 id=&quot;github-agentic-workflows-reach-ga-with-a-defense-in-depth-execution-model&quot; tabindex=&quot;-1&quot;&gt;GitHub Agentic Workflows Reach GA with a Defense-in-Depth Execution Model&lt;/h2&gt;
&lt;p&gt;GitHub Agentic Workflows, an open source project from GitHub Next, became generally available at Universe on October 28 for GitHub Free, Team and Enterprise. Workflows are written as Markdown in natural language and compiled into standard GitHub Actions YAML, so they reuse existing runners and policies. Typical uses are issue triage, CI failure analysis, documentation updates and test writing.&lt;/p&gt;
&lt;p&gt;The execution model is the interesting part for anyone designing agent pipelines. Workflows run with read-only permissions by default, agents execute in a sandboxed container behind an Agent Workflow Firewall, access to GitHub content passes through an integrity filter, and agent outputs go through &amp;quot;safe outputs&amp;quot; validation plus a dedicated threat-detection job before any proposed change takes effect. New at GA, agents can run in Cloud Hypervisor or NVX micro-VMs, and enclaves help prevent leaks in cross-repository workflows.&lt;/p&gt;
&lt;p&gt;The default engine is Copilot CLI (requiring a Copilot license), but teams can bring their own coding agent, choose providers per workflow, or use a model router that picks models by task difficulty. Self-hosted Actions Runner Controller runners are now supported, including Docker-in-Docker, and &lt;code&gt;gh aw forecast&lt;/code&gt; (now stable) estimates monthly AI credit spend including threat-detection costs. The pattern of &amp;quot;natural-language intent, compiled to a reviewable deterministic artifact, executed with least privilege and gated outputs&amp;quot; is becoming the template for CI-hosted agents.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;anthropic-moves-its-monorepo-from-git-to-scion-to-keep-up-with-agent-commit-volume&quot; tabindex=&quot;-1&quot;&gt;Anthropic Moves Its Monorepo from Git to Scion to Keep Up with Agent Commit Volume&lt;/h2&gt;
&lt;p&gt;InfoQ covered Anthropic&#39;s account of replacing Git for its monorepo with Scion, an in-house source control system built for AI-driven development. The driver was volume: daily merged changes rose from about 2,500 in May to more than 30,000 in September, CI performs around 25 million checkouts per day, and the repository holds nearly 100 GB of main-branch history. Shallow clones, sparse checkouts, promisor remotes, snapshots and several hundred Git mirrors had all become harder to operate.&lt;/p&gt;
&lt;p&gt;Scion moves commits, merges, rebases, history and blame to the server. Changesets record files changed relative to the first parent with content-addressed blob pointers, clients fetch immutable trees and blobs directly instead of negotiating packfiles, a node-local daemon shares cached objects between jobs, and regional caches sit in front of authoritative servers backed by Spanner and cloud storage. Checkouts are virtualized and tools can read diffs or perform some writes through APIs without checking out code. Anthropic reports cache hit rates above 99%, a footprint about 200 times smaller than its Git mirrors, and checkouts roughly two orders of magnitude faster, and credits Meta&#39;s Mononoke and Bonsai model for many of the ideas.&lt;/p&gt;
&lt;p&gt;Scion is not a general Git replacement: it depends on Anthropic&#39;s bespoke code review, search and CI, and gives up offline independence. The transferable lessons are about where work happens when agents generate most commits: expose rich server-side repository APIs so agents do not need full checkouts, share immutable objects through caches, and treat checkout cost as a first-class CI bottleneck.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/anthropic-source-control/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;ibm-granite-switch-4.1.1-treats-model-capabilities-as-versioned%2C-swappable-adapters&quot; tabindex=&quot;-1&quot;&gt;IBM Granite Switch 4.1.1 Treats Model Capabilities as Versioned, Swappable Adapters&lt;/h2&gt;
&lt;p&gt;IBM&#39;s Granite Switch 4.1.1 is a patch release of a modular architecture in which a single 30B checkpoint carries built-in adapters that are switched on as needed, so one endpoint can serve chat, decision-making and speech tasks. The release is explicitly versioned like software: 4.1.1 is backwards compatible with 4.1, and the only retrained components are a new Granite Speech ASR model (10 languages, with punctuation and capitalization in the output) and an updated Decision adapter. The base model was not touched.&lt;/p&gt;
&lt;p&gt;The idea is that capabilities can be patched independently, the way a library ships a point release, rather than retraining or redeploying an entire model. IBM says an Adapter SDK is coming that will let developers and third parties build, version and test their own adapters. For teams running self-hosted models, this is a different operational model from swapping whole checkpoints: smaller, more frequent updates with clearer compatibility guarantees.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/ibm-granite/granite-switch-4-1-1&quot;&gt;Read more&lt;/a&gt; — Hugging Face&lt;/p&gt;
&lt;hr /&gt;
&lt;h3 id=&quot;safe-%26-secure-ai-agent-practices&quot; tabindex=&quot;-1&quot;&gt;Safe &amp;amp; Secure AI Agent Practices&lt;/h3&gt;
&lt;h2 id=&quot;docker&#39;s-%22the-cve-that-did-not-exist%22%3A-sandboxes-contain-actions%2C-not-claims&quot; tabindex=&quot;-1&quot;&gt;Docker&#39;s &amp;quot;The CVE That Did Not Exist&amp;quot;: Sandboxes Contain Actions, Not Claims&lt;/h2&gt;
&lt;p&gt;The latest entry in Docker&#39;s Coding Agent Horror Stories series describes a self-hosted AI code review bot, running local models via Ollama in GitLab CI with a read-only rule, that reviewed a routine PostgreSQL patch bump from 16.11 to 16.13. It returned a 1,500-word rejection describing an &amp;quot;architectural takeover,&amp;quot; citing a CVE number that was never issued, naming a company that does not exist, referencing Kubernetes resources it could not have seen, and recommending that the team revoke all IAM access and run Go scripts to isolate &amp;quot;infected&amp;quot; database nodes. No permission was violated; the danger was entirely in a confident, well-formatted report that a tired on-call engineer might act on.&lt;/p&gt;
&lt;p&gt;Docker&#39;s hardened design runs the bot in a Docker Sandbox microVM with the repository mounted read-only, no GitLab token, kubeconfig or cloud credentials, and a deny-by-default network policy that opens exactly one path to the model endpoint. Output is written to a file on the runner rather than posted directly, and a post-processing step checks every CVE identifier in the review against the NVD API, flagging the review as unverified if any ID does not exist. The policy log is reviewed afterward to prove the agent could not reach the cluster.&lt;/p&gt;
&lt;p&gt;The recommendations generalize: treat agent security findings as hypotheses, automatically verify CVE IDs against a public registry, never run agent-written remediation without review, make read-only a property of the environment rather than a convention, and label agent output so it does not borrow the authority of a senior engineer. As the post puts it, containment handles what an agent can do; how much to trust what it says remains a human decision.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.docker.com/blog/coding-agent-horror-stories-the-cve-that-did-not-exist/&quot;&gt;Read more&lt;/a&gt; — Docker&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;google&#39;s-pagebreak-only-reports-vulnerabilities-a-deterministic-validator-can-reproduce&quot; tabindex=&quot;-1&quot;&gt;Google&#39;s PageBreak Only Reports Vulnerabilities a Deterministic Validator Can Reproduce&lt;/h2&gt;
&lt;p&gt;Google disclosed PageBreak, a Gemini-based security agent that has found more than 500 cross-site scripting vulnerabilities in Google&#39;s own web applications since entering full operation in January 2026. The design addresses the same problem as the Docker story above. The model is free to propose any hypothesis, but a finding only reaches a product team once a deterministic validator, plain code written by security engineers and kept outside the agent loop, reproduces it with a real payload: executing JavaScript for XSS, writing to a world-readable location for path traversal, or triggering outbound DNS for remote code execution. Google reports a near-zero false-positive rate.&lt;/p&gt;
&lt;p&gt;Findings the validators cannot confirm are kept internally to seed future scans and to highlight where new validators are needed, rather than being sent to engineers as noise. Google acknowledges the trade-off: validators do not yet cover all vulnerability classes, so false negatives are possible, and the 500+ figure has not been independently verified. PageBreak is being integrated with CodeMender, Google&#39;s agentic fix generator, with the goal that product teams only review AI-generated fixes.&lt;/p&gt;
&lt;p&gt;The pattern, &amp;quot;LLM generates hypotheses, deterministic code decides what counts,&amp;quot; is a practical answer to hallucinated security reports and is applicable well beyond XSS scanning.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/google-pagebreak-security-agent/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;gemini-enterprise-semantic-governance-now-evaluates-tool-calls-from-non-gemini-models&quot; tabindex=&quot;-1&quot;&gt;Gemini Enterprise Semantic Governance Now Evaluates Tool Calls from Non-Gemini Models&lt;/h2&gt;
&lt;p&gt;On October 26, Google extended semantic governance policies on Gemini Enterprise Agent Platform to evaluate tool calls produced by non-Gemini providers, including Anthropic, OpenAI Chat Completions, Mistral and self-hosted OpenAI-compatible servers. Organizations running mixed-model agent fleets can now apply one authorization policy layer regardless of which model proposes a tool call. Streamed responses are not yet supported, and each authorization policy&#39;s &lt;code&gt;when&lt;/code&gt; condition must be updated to match the provider&#39;s request path.&lt;/p&gt;
&lt;p&gt;The same update documents user intent matching, which checks that tool calls align with what the user asked for and is on by default. It can be disabled with the &lt;code&gt;sgpUserIntentMatching&lt;/code&gt; key, but Google notes that doing so reduces protection against indirect prompt injection. Combined with the human-in-the-loop confirmation policies that shipped with the computer use tool&#39;s GA on October 30, this is a concrete example of moving agent guardrails out of prompts and into platform-enforced policy.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available&quot;&gt;GitHub Agentic Workflows generally available&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/anthropic-source-control/&quot;&gt;How Anthropic Redesigned Source Control for AI-Driven Development&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/ibm-granite/granite-switch-4-1-1&quot;&gt;Granite Switch 4.1.1 — Update models like you update software&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.docker.com/blog/coding-agent-horror-stories-the-cve-that-did-not-exist/&quot;&gt;Coding Agent Horror Stories: The CVE That Did Not Exist&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/google-pagebreak-security-agent/&quot;&gt;Google Discloses PageBreak AI Security Agent That Found 500+ XSS Bugs&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes&quot;&gt;Gemini Enterprise Agent Platform release notes&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>GitHub Agentic Workflows reached GA, compiling natural-language Markdown automations into Actions YAML that run read-only by default in a firewalled sandbox, with safe-output validation and threat detection before any change lands.</summary>
    
  </entry>
  
  <entry>
    <title>Spring Ecosystem Update: Spring Boot 4.2.0-RC1, Spring AI 2.1 RC1 Adds Session Memory and Tool Confirmation, Spring Cloud Paddington RC1, 2026-10-25</title>
    <link href="https://javarubberduck.com/java/news-2026-10-25-spring/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-10-25-spring/</id>
    <content type="html">&lt;h2 id=&quot;spring-boot-4.2.0-rc1-moves-buildpacks-to-ubuntu-26.04-and-speeds-up-executable-jar-startup&quot; tabindex=&quot;-1&quot;&gt;Spring Boot 4.2.0-RC1 Moves Buildpacks to Ubuntu 26.04 and Speeds Up Executable Jar Startup&lt;/h2&gt;
&lt;p&gt;October&#39;s Patch Thursday release train brought &lt;strong&gt;Spring Boot 4.2.0-RC1&lt;/strong&gt;, the last milestone before 4.2 goes GA. It contains 113 enhancements, documentation improvements and bug fixes. The default Cloud Native Buildpacks builder is now &lt;code&gt;paketobuildpacks/ubuntu-resolute-builder&lt;/code&gt;, based on &lt;strong&gt;Ubuntu 26.04&lt;/strong&gt;. Teams that build OCI images with &lt;code&gt;bootBuildImage&lt;/code&gt; or &lt;code&gt;spring-boot:build-image&lt;/code&gt; will get the new base image without changing anything, so check any native dependencies or OS-level assumptions in your images before 4.2 ships.&lt;/p&gt;
&lt;p&gt;Two other changes in the RC are worth testing. &lt;code&gt;spring-boot-properties-migrator&lt;/code&gt; can now &lt;strong&gt;fail the application&lt;/strong&gt; when it finds errors or warnings, instead of only logging them. That makes it practical to run in CI and catch renamed or removed properties before an upgrade reaches production. Startup is also faster, including quicker class and resource lookup inside executable jars.&lt;/p&gt;
&lt;p&gt;The same day saw maintenance releases on the supported lines. &lt;strong&gt;Spring Boot 4.1.2&lt;/strong&gt; carries 126 bug fixes, documentation improvements and dependency upgrades, and &lt;strong&gt;Spring Boot 4.0.9&lt;/strong&gt; shipped alongside it. If you are already on 4.x, upgrade to the patch release for your line now, and use the RC to test the 4.2 migration.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-boot-4-2-0-RC1-available-now&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;spring-ai-2.1.0-rc1%3A-session-memory%2C-tool-call-confirmation-and-message-parts-across-providers&quot; tabindex=&quot;-1&quot;&gt;Spring AI 2.1.0-RC1: Session Memory, Tool-Call Confirmation and Message Parts Across Providers&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Spring AI 2.1.0-RC1&lt;/strong&gt; is built on Boot 4.2.0-RC1 and finishes the move to ordered message parts. OpenAI Chat Completions, Anthropic, Google GenAI, Bedrock Converse, Ollama, Mistral AI and DeepSeek now produce and consume &lt;code&gt;MessagePart&lt;/code&gt; content natively. Reasoning blocks, tool calls and opaque payloads such as Anthropic thinking signatures and Gemini thought signatures keep their order across a conversation. A new &lt;code&gt;replay-reasoning-content&lt;/code&gt; option lets you stop replaying &lt;code&gt;reasoning_content&lt;/code&gt; to endpoints such as Groq that reject the field.&lt;/p&gt;
&lt;p&gt;The release adds two features for building agents:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Session memory&lt;/strong&gt;: an append-only event log ported from the community Spring AI Session project. Compaction is turn-aware, so it never separates a tool call from its result. It comes with &lt;code&gt;TurnCountTrigger&lt;/code&gt;, &lt;code&gt;WindowCompactionStrategy&lt;/code&gt; and an optional &lt;code&gt;ChatModelSummarizer&lt;/code&gt;, plus in-memory and JDBC repositories for PostgreSQL, MySQL, MariaDB and H2. Every call must pass a session ID, and there is no shared default session.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Tool-call confirmation&lt;/strong&gt;: &lt;code&gt;@Tool(requiresConfirmation = true)&lt;/code&gt; pauses the tool-calling loop so your application can approve or deny each call through &lt;code&gt;ToolCallDecision&lt;/code&gt;. Policies are set with &lt;code&gt;ToolCallConfirmationPolicy&lt;/code&gt; and &lt;code&gt;spring.ai.tools.confirmation.*&lt;/code&gt; properties.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Search with your own query vector&lt;/strong&gt;: &lt;code&gt;similaritySearch(float[], SearchRequest)&lt;/code&gt; works on pgvector, Qdrant, Redis, Elasticsearch and OpenSearch. Stores that don&#39;t support it fail fast.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Qxotic jinfer&lt;/strong&gt;: a pure-JVM engine for chat, embedding, text-to-speech and transcription models.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Watch for these upgrade issues. &lt;code&gt;ChatMemory&lt;/code&gt;, &lt;code&gt;MessageWindowChatMemory&lt;/code&gt;, &lt;code&gt;MessageChatMemoryAdvisor&lt;/code&gt; and several memory repositories are now &lt;strong&gt;deprecated for removal in 3.0.0&lt;/strong&gt;. &lt;code&gt;BedrockChatOptions.requestParameters&lt;/code&gt; is now a &lt;code&gt;Map&amp;lt;String, Object&amp;gt;&lt;/code&gt;, so code that assigns it to a string-valued map won&#39;t compile. &lt;code&gt;PgVectorStore&lt;/code&gt; no longer writes &lt;code&gt;distance&lt;/code&gt; metadata, so read &lt;code&gt;Document#getScore()&lt;/code&gt; instead. Dependencies move to MCP Java SDK 2.1.0, OpenAI Java SDK 4.80.0 and Anthropic Java SDK 2.72.0. &lt;strong&gt;Spring AI 2.0.2&lt;/strong&gt; ships the same fixes for the 2.0 line.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-ai-2-1-0-RC1-available-now&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;spring-cloud-2026.0.0-rc1-paddington%3A-oauth-2-for-config-clients%2C-websockets-in-gateway-mvc%2C-bus-folded-into-stream&quot; tabindex=&quot;-1&quot;&gt;Spring Cloud 2026.0.0-RC1 Paddington: OAuth 2 for Config Clients, WebSockets in Gateway MVC, Bus Folded Into Stream&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Spring Cloud 2026.0.0-RC1&lt;/strong&gt; (Paddington) moves its modules to 5.1.0-RC1 on Spring Boot 4.2.0-RC1. The most useful change for many teams is in &lt;strong&gt;Spring Cloud Config&lt;/strong&gt;: clients can now authenticate to the Config Server with &lt;strong&gt;OAuth 2&lt;/strong&gt;, not only basic auth or custom headers. &lt;strong&gt;Spring Cloud Gateway MVC&lt;/strong&gt;, the servlet-based gateway, adds &lt;strong&gt;WebSocket support&lt;/strong&gt;, closing one of the remaining gaps with the reactive gateway.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Spring Cloud Kubernetes&lt;/strong&gt; gives the Configuration Watcher high-availability support and upgrades to Fabric8 8.0.0 and Kubernetes Java Client 27.0.0. The release also restructures the project list: &lt;strong&gt;Spring Cloud Bus is no longer a standalone project&lt;/strong&gt; and its code now lives in Spring Cloud Stream. The Bus artifact coordinates haven&#39;t changed, so existing dependencies should resolve without edits.&lt;/p&gt;
&lt;p&gt;The other modules (Circuitbreaker, Commons, Consul, Function, Netflix, OpenFeign, Task, Vault, Zookeeper) only got version bumps in this RC. The Northfields line received a maintenance release, &lt;strong&gt;Spring Cloud 2025.1.4&lt;/strong&gt;, the same day.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-cloud-2026-0-0-RC1-has-been-released&quot;&gt;Read more&lt;/a&gt; — Spring Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-boot-4-2-0-RC1-available-now&quot;&gt;Spring Boot 4.2.0-RC1 available now&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-boot-4-1-2-available-now&quot;&gt;Spring Boot 4.1.2 available now&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-ai-2-1-0-RC1-available-now&quot;&gt;Spring AI 2.1.0-RC1 and 2.0.2 Available Now&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-cloud-2026-0-0-RC1-has-been-released&quot;&gt;Spring Cloud 2026.0.0-RC1 (aka Paddington) Has Been Released&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://spring.io/blog/2026/10/23/spring-cloud-2025-1-4-has-been-released&quot;&gt;Spring Cloud 2025.1.4 (aka Northfields) Has Been Released&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Spring AI 2.1.0-RC1 adds an append-only session memory with turn-aware compaction and human approval of tool calls, on top of the new Spring Boot 4.2.0-RC1.</summary>
    
  </entry>
  
  <entry>
    <title>JavaRubberDuck Podcast — October 25, 2026</title>
    <link href="https://javarubberduck.com/podcast/2026-10-25/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/podcast/2026-10-25/</id>
    <content type="html">&lt;h2 id=&quot;in-this-episode&quot; tabindex=&quot;-1&quot;&gt;In This Episode&lt;/h2&gt;
&lt;p&gt;Stories covered from 2026-10-25 to 2026-10-25:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Anthropic Donates Agent Skills to the Agentic AI Foundation, Joining MCP Under the Linux Foundation&lt;/strong&gt; — Anthropic (&lt;a href=&quot;https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Harvard–Jellyfish Study: AI Agents Lift Code Output 20–30% but Resolved Work Only 12%, as Review Time Rises 49%&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/jellyfish-ai-productivity/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;ThinkingBox Training: Turning Database-State Grading Into RL Rewards Lifts pass@1 From 51.7% to 60.8%&lt;/strong&gt; — Hugging Face Blog (&lt;a href=&quot;https://huggingface.co/blog/microsoft/thinkingbox-training&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;METR and Redwood Red-Team OpenAI&#39;s Internal Agent Monitor: Coverage Gaps, Header Bypasses and Five Design Ideals&lt;/strong&gt; — METR (&lt;a href=&quot;https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;METR&#39;s Review of Anthropic&#39;s Risk Report: Low Risk Comes From Limited Capability, Not Proven Safeguards&lt;/strong&gt; — METR (&lt;a href=&quot;https://metr.org/blog/2026-10-20-misalignment-risk-report-august-2026-review/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Docker Sandboxes + BAND: Multi-Agent Teams With Outbound-Only Connectivity and Proxy-Injected Credentials&lt;/strong&gt; — Docker Blog (&lt;a href=&quot;https://www.docker.com/blog/band-ai-agents-docker-sandboxes/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Claude Code 2.1.303–2.1.310: Safety Telemetry, Code Review Levels and Sandbox Hardening&lt;/strong&gt; — Claude Code Changelog (&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Codex CLI 0.164: Prompt Presets, Enterprise MCP Auth and Credential Masking&lt;/strong&gt; — OpenAI Codex Changelog (&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Copilot Model Policy API Goes GA as Six Models Retire&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-19-copilot-model-policy-api-for-enterprise-orgs-and-teams&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Ollama Adds Text-to-Speech With Gradium&#39;s Phonon and an OpenAI-Compatible Speech Endpoint&lt;/strong&gt; — Ollama Blog (&lt;a href=&quot;https://ollama.com/blog/speech-models&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hermes Agent v0.22.3–v0.22.5: MCP Progress Notifications, Pluggable Delegation and OpenClaw Migration&lt;/strong&gt; — GitHub (NousResearch) (&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;listen&quot; tabindex=&quot;-1&quot;&gt;Listen&lt;/h2&gt;
&lt;audio controls=&quot;&quot;&gt;
  &lt;source src=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-10-25-podcast.mp3&quot; type=&quot;audio/mpeg&quot; /&gt;
  &lt;a href=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-10-25-podcast.mp3&quot;&gt;Download episode&lt;/a&gt;
&lt;/audio&gt;
&lt;blockquote&gt;
&lt;p&gt;🤖 &lt;strong&gt;AI-generated episode.&lt;/strong&gt; Both hosts are synthetic voices produced with Google NotebookLM — no human narrators were recorded. The conversation is generated from the news digest articles linked above.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;subscribe&quot; tabindex=&quot;-1&quot;&gt;Subscribe&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://javarubberduck.com/podcast.xml&quot;&gt;RSS feed&lt;/a&gt; — paste into any podcast app&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://open.spotify.com/show/&quot;&gt;Spotify&lt;/a&gt; — coming soon&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Agent Skills move to the Agentic AI Foundation, a Harvard study finds AI agents shift the bottleneck to code review, and METR red-teams OpenAI&#39;s internal agent monitoring.</summary>
    
  </entry>
  
  <entry>
    <title>Java News: Apache Camel Patches 38 CVEs Found by Claude, TornadoVM 7.3 Adds cuTile and Multi-GPU, Babylon HAT Brings GPU Kernels to Java, 2026-10-25</title>
    <link href="https://javarubberduck.com/java/news-2026-10-25-java-jdk/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-10-25-java-jdk/</id>
    <content type="html">&lt;h2 id=&quot;java-roundup%3A-apache-camel-fixes-38-cves%2C-payara-closes-two-rces%2C-gatherers4j-0.15-and-adk-for-java-1.12&quot; tabindex=&quot;-1&quot;&gt;Java Roundup: Apache Camel Fixes 38 CVEs, Payara Closes Two RCEs, Gatherers4j 0.15 and ADK for Java 1.12&lt;/h2&gt;
&lt;p&gt;InfoQ&#39;s roundup for the week of October 12 had more security news than usual. Apache Camel published advisories for &lt;strong&gt;38 CVEs&lt;/strong&gt;: 13 rated HIGH and 25 rated MEDIUM. Claude found 28 of them through Anthropic&#39;s Project Glasswing, and outside security researchers reported the other 10. The fixed releases are &lt;strong&gt;Camel 4.23.0, 4.22.1 and 4.18.5&lt;/strong&gt;, depending on which release line you run. Camel sits inside a lot of integration middleware, so even teams that never added it directly should check their transitive dependencies.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Azul Payara 7.5.0&lt;/strong&gt; (the October 2026 edition) ships as Community Edition 7.2026.10 and Enterprise Editions 6.43.0 and 5.92.0. It fixes two remote code execution flaws. CVE-2026-2586 can be exploited through the Administration Console to run OS commands as the service user. CVE-2026-2587 is reached through server-side template rendering and could lead to host compromise. The release also lets SSH nodes run without elevated permissions and adds &lt;code&gt;delete-jvm-option&lt;/code&gt; and &lt;code&gt;delete-system-properties&lt;/code&gt; for command-line configuration.&lt;/p&gt;
&lt;p&gt;On the library side, &lt;strong&gt;Gatherers4j 0.15.0&lt;/strong&gt; adds stream gatherers for statistics: &lt;code&gt;movingStatistics()&lt;/code&gt;, &lt;code&gt;runningStatistics()&lt;/code&gt; and their &lt;code&gt;By&lt;/code&gt; variants each emit a &lt;code&gt;StatisticalSummary&lt;/code&gt; record per element. It also adds &lt;code&gt;movingWeightedMean()&lt;/code&gt; with linear or custom weights. &lt;strong&gt;Google ADK for Java 1.12.0&lt;/strong&gt; adds workflow graph types (the &lt;code&gt;Node&lt;/code&gt; interface, &lt;code&gt;BaseNode&lt;/code&gt;, &lt;code&gt;Start&lt;/code&gt; and an &lt;code&gt;Edge&lt;/code&gt; record) and deprecates &lt;code&gt;maxInvocations&lt;/code&gt; in &lt;code&gt;ContextCacheConfig&lt;/code&gt;. ADK for Kotlin 1.4.0 adds an &lt;code&gt;EdgesDsl&lt;/code&gt; and a &lt;code&gt;ChatCompletions&lt;/code&gt; backend for OpenAI-compatible endpoints. &lt;strong&gt;Yupiik Fusion 1.2.0&lt;/strong&gt; has a breaking change: &lt;code&gt;HandlebarsCompiler&lt;/code&gt; now behaves more like Handlebars.js.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct12-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;tornadovm-7.3-adds-a-cutile-cuda-backend%2C-nvidia-library-calls-and-multi-gpu-execution&quot; tabindex=&quot;-1&quot;&gt;TornadoVM 7.3 Adds a cuTile CUDA Backend, NVIDIA Library Calls and Multi-GPU Execution&lt;/h2&gt;
&lt;p&gt;TornadoVM, the open-source framework that offloads Java code to GPUs, has reached &lt;strong&gt;version 7.3&lt;/strong&gt;. It runs on &lt;strong&gt;JDK 21, 25, 26 and 27&lt;/strong&gt; next to a standard JVM (OpenJDK, GraalVM, Temurin or Zulu), so you don&#39;t need a custom runtime. The 7.x line finishes an architectural cleanup: backends use the Foreign Function &amp;amp; Memory API instead of JNI, the experimental JVMCI dependency is gone, and compiler metadata goes through a single reflection, ASM and Unsafe layer on every supported JDK.&lt;/p&gt;
&lt;p&gt;The new backend work targets NVIDIA hardware. An optimized CUDA backend supports the &lt;strong&gt;cuTile&lt;/strong&gt; programming model through a new &lt;code&gt;TileContext&lt;/code&gt;. TaskGraphs can now call CUDA libraries such as &lt;strong&gt;cuBLAS and CUTLASS&lt;/strong&gt; directly, so you can swap a JIT-generated kernel for a tuned library call, chain several calls, or mix both in one graph. The hybrid API gives the same library access for Apple&#39;s MXL. This release also adds &lt;strong&gt;multi-GPU execution&lt;/strong&gt;. OpenCL and Apple Metal remain compilation targets for Java source.&lt;/p&gt;
&lt;p&gt;TornadoVM now has four programming styles: the loop-parallel annotation API, a kernel-context API for finer control, &lt;code&gt;TileContext&lt;/code&gt;, and the hybrid library API. Integrations include Quarkus, Flink and Spring AI. In the Flink integration, you mark a Table API query with a configuration flag, and at execution time a library provider inspects the query&#39;s intermediate representation and decides whether to run it on the GPU. The team also released &lt;strong&gt;jitLLM 2.0&lt;/strong&gt;, a Java inference engine built on TornadoVM that ships as a single JAR and is aimed at the same use cases as llama.cpp and Ollama.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/tornadovm-7-gpu/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;reflect%2C-transform%2C-accelerate%3A-project-babylon&#39;s-hat-compiles-java-code-models-to-cuda-and-opencl&quot; tabindex=&quot;-1&quot;&gt;Reflect, Transform, Accelerate: Project Babylon&#39;s HAT Compiles Java Code Models to CUDA and OpenCL&lt;/h2&gt;
&lt;p&gt;In a Devoxx Belgium 2026 talk published on Inside Java, Juan Fumero walked through the &lt;strong&gt;Heterogeneous Accelerator Toolkit (HAT)&lt;/strong&gt;. HAT is a parallel programming framework built on &lt;strong&gt;code reflection&lt;/strong&gt; from OpenJDK Project Babylon. Where TornadoVM works on bytecode, HAT uses Babylon&#39;s code reflection API to turn annotated Java compute code into a code model that can be analyzed and rewritten.&lt;/p&gt;
&lt;p&gt;HAT works in three stages. First it reflects the Java method into a code model. Then it transforms that model into a GPU-friendly form. Finally it offloads the selected computations to &lt;strong&gt;OpenCL or CUDA&lt;/strong&gt;. The talk&#39;s examples are tensor-core programming and matrix multiplication, the operations behind most AI and NLP inference. It also shows how to profile and tune the generated CUDA kernels with &lt;strong&gt;NVIDIA Nsight Compute and Nsight Systems&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;Babylon is still an OpenJDK research project, not a targeted JEP, so HAT is not something to adopt in production yet. The talk is a good introduction to code reflection for developers who haven&#39;t used it. Code reflection is the mechanism Babylon proposes for running Java on GPUs, translating it to SQL, or differentiating it, all without leaving the language.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://inside.java/2026/10/20/devoxxbelgium-bringing-gpu-computing-to-java/&quot;&gt;Read more&lt;/a&gt; — Inside Java&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct12-2026/&quot;&gt;Java News Roundup: JEPs for JDK 28, Azul Payara, Gatherers4j, Google ADK, Yupiik, Apache Camel CVEs&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/tornadovm-7-gpu/&quot;&gt;TornadoVM 7 Enables GPU-Native Workloads from Java Applications&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://inside.java/2026/10/20/devoxxbelgium-bringing-gpu-computing-to-java/&quot;&gt;Reflect, Transform, Accelerate: Bringing GPU Computing to Java&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Apache Camel fixed 38 CVEs, 28 of them found by Claude through Project Glasswing, while Azul Payara 7.5.0 closed two remote code execution flaws.</summary>
    
  </entry>
  
  <entry>
    <title>Developer Tools Digest: Claude Code Adds Safety Telemetry and Review Levels, Codex CLI 0.164 Presets, Ollama Ships Speech Models, 2026-10-25</title>
    <link href="https://javarubberduck.com/ai/news-2026-10-25-dev-tools/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-10-25-dev-tools/</id>
    <content type="html">&lt;h2 id=&quot;claude-code-2.1.303%E2%80%932.1.310%3A-safety-telemetry%2C-code-review-levels-and-sandbox-hardening&quot; tabindex=&quot;-1&quot;&gt;Claude Code 2.1.303–2.1.310: Safety Telemetry, Code Review Levels and Sandbox Hardening&lt;/h2&gt;
&lt;p&gt;Claude Code shipped eight releases between October 18 and 23. The new features fall into three groups: observability, review and setup.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;&lt;code&gt;claude_code.safety_event&lt;/code&gt;&lt;/strong&gt; (2.1.307): an OpenTelemetry event for each tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back. Teams can now chart how often their guardrails fire.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Code Review levels&lt;/strong&gt; (2.1.304): Quick, Standard (the default) or Deep, set per organization or per repository. &lt;code&gt;@claude review stop&lt;/code&gt; ends an &lt;code&gt;@claude review always&lt;/code&gt; subscription on a pull request.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;&lt;code&gt;/auto-mode-setup&lt;/code&gt;&lt;/strong&gt; (2.1.304): a guided setup that reads your project and recent sessions to teach auto mode about your environment, now also on Bedrock, Vertex and Foundry.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;&lt;code&gt;/claude-api&lt;/code&gt; additions&lt;/strong&gt;: &lt;code&gt;managed-agents-secure&lt;/code&gt; compares a Managed Agents configuration against Anthropic&#39;s security guide, and &lt;code&gt;api-version&lt;/code&gt; migrates code to API version 2026-10-21.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Several changes tighten the sandbox and the permission system. In 2.1.309, sandboxed Bash on macOS and Linux no longer runs shell startup files outside the sandbox before each command. Sandboxed commands and auto-accepted edits can no longer modify the file passed with &lt;code&gt;--mcp-config&lt;/code&gt;. PowerShell hooks that exit with code 2 now block correctly. Dangerous &lt;code&gt;rm&lt;/code&gt;/&lt;code&gt;rmdir&lt;/code&gt; commands are denied immediately in bypass-permissions mode in Desktop and VS Code, and &lt;code&gt;claude ultrareview&lt;/code&gt; now requires workspace trust. In 2.1.305, the new &lt;code&gt;sandbox.allowNestedUserNamespaces&lt;/code&gt; setting lets sandboxed commands on Linux create user namespaces (for example Chromium or rootless podman). The changelog warns that this can weaken file protections, so enable it only when you need it.&lt;/p&gt;
&lt;p&gt;A batch of fixes targets &lt;strong&gt;LLM gateways&lt;/strong&gt;. When a gateway rejects tool search, &lt;code&gt;context_management&lt;/code&gt; fields, the effort beta value or mid-conversation system messages, Claude Code now retries without them. HTTP MCP tool calls are no longer sent twice when the connection drops; only read-only or idempotent tools are retried. The Monitor tool now works on Bedrock, Vertex and Foundry (2.1.310).&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Read more&lt;/a&gt; — Claude Code Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;codex-cli-0.164%3A-prompt-presets%2C-enterprise-mcp-auth-and-credential-masking&quot; tabindex=&quot;-1&quot;&gt;Codex CLI 0.164: Prompt Presets, Enterprise MCP Auth and Credential Masking&lt;/h2&gt;
&lt;p&gt;OpenAI released &lt;strong&gt;Codex CLI 0.164.0&lt;/strong&gt; on October 23. The main addition is &lt;code&gt;/presets&lt;/code&gt;, which gives you ten numbered slots for saved prompts and slash commands behind a keyboard menu. &lt;code&gt;/model&lt;/code&gt; now accepts inline arguments for model, reasoning effort and speed, and &lt;code&gt;/permissions&lt;/code&gt; accepts a permission profile name. Together these let you switch configurations with one command. Kitty and Sixel terminals now show image previews in the fullscreen transcript, and the Amazon Bedrock setup wizard is on by default.&lt;/p&gt;
&lt;p&gt;For enterprise use, MCP servers can now authenticate through eligible Codex account grants or a dedicated sign-in, including OAuth clients that need a client secret. &lt;strong&gt;Credential masking&lt;/strong&gt; now covers remote Bash commands and Zsh snapshot sessions, including credentials from custom providers, so secrets stop appearing in transcripts and logs.&lt;/p&gt;
&lt;p&gt;The release also has reliability and security fixes. A failed compaction no longer leaves a partial summary in history. Resumed sessions find compressed histories and restore saved thread settings. Windows proxy authentication uses the signed-in user&#39;s credentials. Sandbox checks were hardened against case-variant macOS paths and replaceable symlinks.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;Read more&lt;/a&gt; — OpenAI Codex Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;github-copilot-model-policy-api-goes-ga-as-six-models-retire&quot; tabindex=&quot;-1&quot;&gt;GitHub Copilot Model Policy API Goes GA as Six Models Retire&lt;/h2&gt;
&lt;p&gt;On October 19, GitHub made the &lt;strong&gt;Copilot model policy REST API&lt;/strong&gt; generally available. GET and PATCH endpoints control which Copilot models are available at four scopes: enterprise (&lt;code&gt;/enterprises/{enterprise}/copilot/model-policies&lt;/code&gt;), enterprise teams, agent apps, and organizations (&lt;code&gt;/orgs/{org}/copilot/model-policies&lt;/code&gt;). An enterprise can keep control itself, delegate it to organizations, or delegate to teams and apps. Platform teams can now manage model access as code instead of clicking through settings pages.&lt;/p&gt;
&lt;p&gt;The timing fits the same day&#39;s retirements. &lt;strong&gt;Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5&lt;/strong&gt; were deprecated. The suggested replacements are Gemini 3.8 Flash, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7. Enterprise admins may need to enable the replacements in their model policies before users see them, and that is a single PATCH call with the new API.&lt;/p&gt;
&lt;p&gt;GitHub also shipped related admin features during the week. The Copilot usage metrics API now reports custom skills, agents and MCP servers, as well as usage from third-party tools such as OpenCode. A global default for Copilot features is GA, and the Copilot impact dashboard adds industry comparisons.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-19-copilot-model-policy-api-for-enterprise-orgs-and-teams&quot;&gt;Read more&lt;/a&gt; — GitHub Changelog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;ollama-adds-text-to-speech-with-gradium&#39;s-phonon-and-an-openai-compatible-speech-endpoint&quot; tabindex=&quot;-1&quot;&gt;Ollama Adds Text-to-Speech With Gradium&#39;s Phonon and an OpenAI-Compatible Speech Endpoint&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Ollama 0.40.11&lt;/strong&gt; adds speech models, starting with &lt;strong&gt;Phonon&lt;/strong&gt;, a family of small text-to-speech models from Gradium. The English &lt;code&gt;phonon&lt;/code&gt; model is Apache 2.0. The &lt;code&gt;phonon:40m-*&lt;/code&gt; variants for French, German, Spanish and Portuguese, plus a five-language multilingual model, are CC BY-NC 4.0, which means noncommercial use only. Speech currently runs on Apple Silicon Macs through Apple&#39;s MLX framework. Windows and Linux support is listed as coming soon.&lt;/p&gt;
&lt;p&gt;From the CLI, &lt;code&gt;ollama run phonon &amp;quot;text&amp;quot;&lt;/code&gt; speaks the text, &lt;code&gt;--voice&lt;/code&gt; picks a voice from Gradium&#39;s catalog, and redirecting output writes a 24 kHz mono WAV file. You can also pipe a chat model&#39;s reply straight into Phonon, and speech starts as soon as the first sentence arrives. For applications, Ollama exposes &lt;code&gt;POST /v1/audio/speech&lt;/code&gt; with OpenAI-compatible &lt;code&gt;model&lt;/code&gt;, &lt;code&gt;voice&lt;/code&gt;, &lt;code&gt;input&lt;/code&gt; and &lt;code&gt;response_format&lt;/code&gt; (WAV or raw PCM) parameters. Existing OpenAI SDK code works if you point its base URL at &lt;code&gt;localhost:11434/v1&lt;/code&gt;, and responses stream, so playback can start right away.&lt;/p&gt;
&lt;p&gt;Ollama reports it ran &lt;strong&gt;35% faster than Gradium&#39;s own runtime&lt;/strong&gt; on an M5 Max MacBook Pro (bf16, 45 seconds of speech, median of 10 runs). Word error rate matched the Hugging Face Transformers reference implementation on 200 Seed-TTS English sentences. Together with the transcription support that local runtimes already have, this lets developers build a fully local voice loop on a laptop.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://ollama.com/blog/speech-models&quot;&gt;Read more&lt;/a&gt; — Ollama Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;hermes-agent-v0.22.3%E2%80%93v0.22.5%3A-mcp-progress-notifications%2C-pluggable-delegation-and-openclaw-migration&quot; tabindex=&quot;-1&quot;&gt;Hermes Agent v0.22.3–v0.22.5: MCP Progress Notifications, Pluggable Delegation and OpenClaw Migration&lt;/h2&gt;
&lt;p&gt;NousResearch shipped four Hermes Agent releases after v0.22.0. &lt;strong&gt;v0.22.2&lt;/strong&gt; (October 15) corrected v0.22.1, which had been built from the wrong commit. It fixes the gateway failing to start after a reboot under systemd, launchd and Docker supervisors, and makes long local-model turns more reliable on laptops. One known issue remains: sessions created under 0.22.1 can&#39;t be deleted.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;v0.22.3&lt;/strong&gt; (October 21) rolls up about 620 merged PRs. The full notes are deferred to v0.23.0, but notable additions include MCP progress notifications shown as live tool progress, a refreshed classic CLI layout, and plugin-declared OAuth clients. v0.22.4 followed the same day with small fixes: plugins that install no dependencies skip the empty consent prompt, and WhatsApp can be safely re-paired after logging out.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;v0.22.5&lt;/strong&gt; (October 22–23, about 81 PRs) reworks Nous sign-in, adds usage-window plan support, and introduces &lt;strong&gt;pluggable delegation backends&lt;/strong&gt;, so subagent work can go to different runtimes. It also adds a &lt;code&gt;hermes import-agent&lt;/code&gt; command and a migration path from &lt;strong&gt;OpenClaw&lt;/strong&gt;, which makes switching between the two personal-assistant projects easier. &lt;code&gt;hermes update&lt;/code&gt; now refuses to offer a held-back stable release.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases&quot;&gt;Read more&lt;/a&gt; — GitHub (NousResearch)&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;Claude Code Changelog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;OpenAI Codex Changelog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-19-copilot-model-policy-api-for-enterprise-orgs-and-teams&quot;&gt;Copilot model policy API for enterprise, orgs, and teams&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.blog/changelog/2026-10-19-github-copilot-models-deprecated-on-october-19-2026&quot;&gt;GitHub Copilot models deprecated on October 19, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://ollama.com/blog/speech-models&quot;&gt;Speech models in Ollama&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases&quot;&gt;Hermes Agent Releases&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Claude Code 2.1.303-2.1.310 added a safety_event OpenTelemetry event, Quick/Standard/Deep Code Review levels and guided auto mode setup, and fixed sandbox bypasses.</summary>
    
  </entry>
  
  <entry>
    <title>Cloud &amp; Infrastructure News: Amazon ECS Gateway, AWS Continuum GA, PHP2Shell Patches Hit Cloud Run, TorchTPU for PyTorch, 2026-10-25</title>
    <link href="https://javarubberduck.com/cloud/news-2026-10-25-cloud/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/cloud/news-2026-10-25-cloud/</id>
    <content type="html">&lt;h2 id=&quot;amazon-ecs-gateway%3A-managed-ingress-that-shares-albs-across-up-to-25-services&quot; tabindex=&quot;-1&quot;&gt;Amazon ECS Gateway: Managed Ingress That Shares ALBs Across Up to 25 Services&lt;/h2&gt;
&lt;p&gt;AWS introduced &lt;strong&gt;Amazon ECS Gateway&lt;/strong&gt;, an account-level resource that provisions and manages ingress for ECS services in your own account: Application Load Balancers, certificates and security groups. It splits the work between two roles. A platform team creates a gateway that sets the shared standards (subnets, security groups, TLS and listener settings, a WAF web ACL) and associates it with clusters. Application teams then opt a service in by adding a &lt;code&gt;--gateway-connectors&lt;/code&gt; block to &lt;code&gt;CreateService&lt;/code&gt; or &lt;code&gt;UpdateService&lt;/code&gt;.&lt;/p&gt;
&lt;p&gt;Routing is host-based, and up to &lt;strong&gt;25 services share one ALB&lt;/strong&gt;. Each service gets its own DNS hostname, listener rule, target group and ACM certificate. ECS adds ALBs as capacity fills, up to a default &lt;code&gt;maxAlbCount&lt;/code&gt; of 10, and removes them when no routes remain. A service&#39;s security group accepts traffic only from the ALB. At deployment, ECS connects the service to every compatible gateway: one with a matching access type (public or VPC-internal), in the same VPC, covering all of the service&#39;s Availability Zones. ECS Gateway works with &lt;code&gt;awsvpc&lt;/code&gt; networking on EC2, Fargate and ECS Managed Instances.&lt;/p&gt;
&lt;p&gt;The gateway itself is free. You pay for the ALBs, certificates, CloudWatch data and data transfer it creates. It is available in every commercial Region that has ECS. Services that already have their own load balancer can add a connector and run both paths while they validate the migration. Note that a gateway&#39;s listener protocol and port can&#39;t be changed after creation, so choose them carefully.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/introducing-amazon-ecs-gateway-a-fully-managed-ingress-solution-for-ecs-services/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;aws-continuum-reaches-ga%3A-exploit-verified-code-vulnerabilities-with-fixes-as-pull-requests&quot; tabindex=&quot;-1&quot;&gt;AWS Continuum Reaches GA: Exploit-Verified Code Vulnerabilities With Fixes as Pull Requests&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;AWS Continuum for code vulnerabilities&lt;/strong&gt; is now generally available in US East (N. Virginia) and US West (Oregon). Continuum builds a context graph from your repositories (through a GitHub integration or an S3 bucket) and, optionally, from an AWS collector running where your code is deployed. It uses that graph to find business-critical systems and their dependencies. Findings are ranked by how close their attack path runs to those systems, not by a generic CVSS-style severity.&lt;/p&gt;
&lt;p&gt;Each finding comes with an evidence level, such as &amp;quot;Static proven&amp;quot;. When Continuum can reproduce an issue, it &lt;strong&gt;runs the exploit in an isolated sandbox&lt;/strong&gt; and attaches proof-of-concept code and reproduction steps, along with an attack path, source analysis and blast-radius assessment. It also checks whether your existing defenses would detect and stop the validated path.&lt;/p&gt;
&lt;p&gt;Fixes arrive as code patches, dependency upgrades or configuration changes, opened as a &amp;quot;Continuum fix&amp;quot; pull request with the risk assessment and proof in the description. A person still has to review and merge each PR, and AWS notes that a generated patch doesn&#39;t by itself mean the issue is fixed. Autonomous application of fixes is described as a future capability, gated on test coverage, scope bounds and historical success rate. The blog post gives no pricing details.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-continuum-for-code-vulnerabilities-is-now-generally-available/&quot;&gt;Read more&lt;/a&gt; — AWS News Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;php2shell%3A-two-critical-php-rces-patched%2C-and-google-cloud-asks-teams-to-rebuild-cloud-run-services&quot; tabindex=&quot;-1&quot;&gt;PHP2Shell: Two Critical PHP RCEs Patched, and Google Cloud Asks Teams to Rebuild Cloud Run Services&lt;/h2&gt;
&lt;p&gt;On October 21, PHP shipped &lt;strong&gt;8.5.12, 8.4.27, 8.3.36 and 8.2.35&lt;/strong&gt; to fix two critical remote code execution flaws nicknamed &amp;quot;php2shell&amp;quot; (CVSS v4.0 9.2). &lt;strong&gt;CVE-2026-109618&lt;/strong&gt; is a memory-safety failure in out-of-memory handling during the request lifecycle. &lt;strong&gt;CVE-2026-109619&lt;/strong&gt; is a use-after-free-style bug in which state outlives the memory backing it. Sansec says both are present in every unpatched release since PHP 5.4. Amazon Threat Intelligence saw an exploitation attempt dated October 17, four days before the patches. RCE is confirmed on Apache mod_php; for PHP-FPM, only the session-lifecycle case has been validated. A third issue, CVE-2026-115560, is still unpatched; set &lt;code&gt;session.upload_progress.enabled=0&lt;/code&gt; until the next bugfix release.&lt;/p&gt;
&lt;p&gt;The cloud fallout continued through the week. On October 22 and 24, Google Cloud published patched runtimes for &lt;strong&gt;Cloud Run, Cloud Run functions, App Engine and Buildpacks&lt;/strong&gt; across Go, Node.js, Java, PHP, Python, Ruby and .NET, plus the google-22 and google-24 stacks. Google&#39;s guidance is to &lt;strong&gt;rebuild and redeploy&lt;/strong&gt; your services with the updated dependencies. Docker published its own advisory for customers. InfoQ reports that official Docker images lagged behind some distributions, and that Cloudflare deployed WAF rules about nine hours before the release.&lt;/p&gt;
&lt;p&gt;Two operational points from the advisories are easy to miss. First, restart every PHP-FPM and Apache worker, because a running worker keeps the old engine in memory after the package is upgraded. Second, check the full distro package version, not the output of &lt;code&gt;php -v&lt;/code&gt;. Distros backport fixes without changing the reported version; Ubuntu 24.04&#39;s fixed package still reports 8.3.6. PHP 8.2 reaches end of life on December 31, 2026.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/php-critical-rce-patches/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;google-cloud%3A-cloud-run-direct-vpc-ingress-preview%2C-firestore-dml-ga%2C-bigquery-ai.translate-ga&quot; tabindex=&quot;-1&quot;&gt;Google Cloud: Cloud Run Direct VPC Ingress Preview, Firestore DML GA, BigQuery AI.TRANSLATE GA&lt;/h2&gt;
&lt;p&gt;Several developer-facing changes appeared in the Google Cloud release notes on October 21–22. &lt;strong&gt;Cloud Run Direct VPC ingress&lt;/strong&gt; entered Preview: instances can use their own ephemeral private IP to communicate with other VPC resources, which makes Cloud Run services look more like ordinary VPC workloads to the rest of the network. &lt;strong&gt;Firestore data modification language (DML)&lt;/strong&gt; operations, including &lt;code&gt;insert&lt;/code&gt; and &lt;code&gt;upsert&lt;/code&gt;, became generally available, which lets you run set-based writes instead of looping over documents in client code.&lt;/p&gt;
&lt;p&gt;In BigQuery, the &lt;strong&gt;&lt;code&gt;AI.TRANSLATE&lt;/code&gt;&lt;/strong&gt; SQL function is now GA. It covers both standard Neural Machine Translation and Translation LLM workloads, so translation runs inside a query instead of a separate pipeline. Dataflow added a Preview of &lt;strong&gt;streaming high availability&lt;/strong&gt; for Streaming Engine, enabled with the &lt;code&gt;requires_high_availability&lt;/code&gt; service option. The CodeMender agent (v0.16.0) in Gemini Enterprise Agent Platform now runs a deep scan by default with &lt;code&gt;cm find&lt;/code&gt;, has a &lt;code&gt;--quick&lt;/code&gt; mode, and adds a &lt;code&gt;cm health&lt;/code&gt; command that checks credentials before a scan.&lt;/p&gt;
&lt;p&gt;Teams on Cloud SQL should also act on the October 23 security bulletin GCP-2026-071: upgrade the Cloud SQL Auth Proxy to 2.26.3 or later and the Proxy Operator to 1.8.4 or later, and update the CA certificates used by language connectors.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/release-notes&quot;&gt;Read more&lt;/a&gt; — Google Cloud Release Notes&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;torchtpu-makes-tpus-a-native-pytorch-device%2C-with-day-one-support-in-transformers-and-diffusers&quot; tabindex=&quot;-1&quot;&gt;TorchTPU Makes TPUs a Native PyTorch Device, With Day-One Support in Transformers and Diffusers&lt;/h2&gt;
&lt;p&gt;Google and Meta open-sourced &lt;strong&gt;TorchTPU&lt;/strong&gt; at PyTorch Conference 2026. It is a native PyTorch backend that makes TPUs a regular device next to CPU and GPU. Before this, PyTorch on TPU meant PyTorch/XLA, with lazy tensors and frequent graph recompilations. With TorchTPU you write ordinary eager PyTorch and add &lt;code&gt;torch.compile&lt;/code&gt; when you want more speed. It includes TPU-tuned Pallas kernels and distributed training support. Install it on a TPU VM with &lt;code&gt;pip install torch-tpu&lt;/code&gt; (add the CPU wheel index so pip doesn&#39;t pull the CUDA build of torch).&lt;/p&gt;
&lt;p&gt;Hugging Face supports it from day one. In &lt;strong&gt;Transformers v5&lt;/strong&gt;, &lt;code&gt;device_map=&amp;quot;tpu&amp;quot;&lt;/code&gt; loads a model straight onto the TPU, and Optimum TPU is no longer needed. In &lt;strong&gt;Diffusers&lt;/strong&gt;, &lt;code&gt;pipe.to(&amp;quot;tpu&amp;quot;)&lt;/code&gt; works, and a &lt;code&gt;TensorParallelConfig&lt;/code&gt; shards large models across chips while they load. The Kernels Hub serves ready-made TPU kernels; a Tokamax ragged-dot kernel plugged into MoE models made decode &lt;strong&gt;21.5× faster on Qwen3-30B-A3B&lt;/strong&gt; and &lt;strong&gt;12.8× faster on Gemma 4 26B-A4B&lt;/strong&gt; at batch size 64 on a v6e-8, compared with the default grouped matmul.&lt;/p&gt;
&lt;p&gt;Other numbers from Hugging Face: Qwen-Image-2.1 at 1024×1024 with 40 steps takes 3.5 s compiled versus 8.0 s eager on a v6e-4 with tensor parallelism of 4. And 96.5% of the Transformers tests that run on TPU pass. Teams that were tied to CUDA because of PyTorch now have a real option to try Cloud TPU capacity without rewriting their models.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/hf-torchtpu&quot;&gt;Read more&lt;/a&gt; — Hugging Face Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;cloudflare-adds-on-demand-cpu-and-heap-profiling-for-production-workers-and-durable-objects&quot; tabindex=&quot;-1&quot;&gt;Cloudflare Adds On-Demand CPU and Heap Profiling for Production Workers and Durable Objects&lt;/h2&gt;
&lt;p&gt;Cloudflare now lets developers &lt;strong&gt;profile Workers and Durable Objects in production&lt;/strong&gt; on demand. A capture attaches to an isolate that is already serving real traffic. For CPU profiles, the runtime starts a V8 sampling profiler at a 1 ms interval and releases the isolate lock while requests continue, then reacquires it to collect the results. Heap-allocation profiles show which functions create objects and add memory pressure. The two views often disagree, since a function can dominate one and not appear in the other.&lt;/p&gt;
&lt;p&gt;Captures start from the Workers Observability dashboard or the CLI. You choose a deployed Worker version or, for Durable Objects, a specific named instance; the request is routed to the machine that hosts that object. Results appear as interactive flamegraphs and can be downloaded in a format pprof can read. For readable TypeScript function names, enable source maps. The target also needs enough traffic during the capture window to produce useful samples.&lt;/p&gt;
&lt;p&gt;Cloudflare&#39;s own examples show the kind of problems this finds. A recursive JSON replacer was rewritten to run 2.7× faster. A duplicate metrics call was using about 1% of CPU. Partially disabled Prometheus instrumentation still held memory, and removing it brought a Worker&#39;s p99.9 memory from 133 MB to 118 MB, back under the 128 MB limit. Captures are short and won&#39;t catch intermittent issues, and Cloudflare says continuous profiling is the next step.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/cloudflare-cpu-memory-profiling/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/introducing-amazon-ecs-gateway-a-fully-managed-ingress-solution-for-ecs-services/&quot;&gt;Introducing Amazon ECS Gateway: A fully-managed ingress solution for ECS services&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/aws/aws-continuum-for-code-vulnerabilities-is-now-generally-available/&quot;&gt;AWS Continuum for code vulnerabilities is now generally available&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/php-critical-rce-patches/&quot;&gt;PHP Patches Two Critical RCE Flaws That Were Already Under Attack&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.docker.com/blog/security-update-on-php2shell-impact-for-docker-customers/&quot;&gt;Security Update on PHP2Shell: Impact for Docker Customers&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://docs.cloud.google.com/release-notes&quot;&gt;Google Cloud Release Notes&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/hf-torchtpu&quot;&gt;Hugging Face ❤️ TorchTPU: TPUs, Natively&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/cloudflare-cpu-memory-profiling/&quot;&gt;Cloudflare Brings On-Demand CPU and Memory Profiling to Production Workers&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Amazon ECS Gateway gives ECS services managed ingress: platform teams define shared ALB, TLS and WAF standards once, and services opt in with a gateway connector.</summary>
    
  </entry>
  
  <entry>
    <title>AI Dev Patterns: Agent Skills Move to the Agentic AI Foundation, Harvard Study Finds AI Coding Gains Stall at Review, ThinkingBox RL Training, 2026-10-25</title>
    <link href="https://javarubberduck.com/ai/news-2026-10-25-ai-patterns/"/>
    <updated>2026-10-25T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/ai/news-2026-10-25-ai-patterns/</id>
    <content type="html">&lt;h2 id=&quot;anthropic-donates-agent-skills-to-the-agentic-ai-foundation%2C-joining-mcp-under-the-linux-foundation&quot; tabindex=&quot;-1&quot;&gt;Anthropic Donates Agent Skills to the Agentic AI Foundation, Joining MCP Under the Linux Foundation&lt;/h2&gt;
&lt;p&gt;On October 22, Anthropic donated the &lt;strong&gt;Agent Skills&lt;/strong&gt; open standard to the &lt;strong&gt;Agentic AI Foundation (AAIF)&lt;/strong&gt;, which the Linux Foundation hosts and which already holds the Model Context Protocol. A skill is a folder containing a &lt;a href=&quot;http://skill.md/&quot;&gt;SKILL.md&lt;/a&gt; file plus whatever scripts and reference material it needs. The donation covers the specification, the name and logo, the GitHub organization and the &lt;a href=&quot;http://agentskills.io/&quot;&gt;agentskills.io&lt;/a&gt; domain. Engineers from Anthropic, Google, OpenAI and other companies will maintain it as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. You don&#39;t have to be an AAIF member to use or contribute to the standard.&lt;/p&gt;
&lt;p&gt;The adoption numbers explain why it needed a neutral home. Since the format was published as an open standard in December 2025, more than &lt;strong&gt;70 agent platforms&lt;/strong&gt; have integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code. About &lt;strong&gt;1 million skills&lt;/strong&gt; have been published and nearly &lt;strong&gt;280 million&lt;/strong&gt; installed. At that scale, one vendor controlling the format would hold back the cross-tool portability that made it useful.&lt;/p&gt;
&lt;p&gt;The proposed priorities matter most to developers: &lt;strong&gt;versioned releases&lt;/strong&gt;, a &lt;strong&gt;formal proposal process&lt;/strong&gt; for spec changes, and &lt;strong&gt;security guidance for skill authors and installers&lt;/strong&gt;. The last one is the most needed. Earlier ecosystem audits found that a large share of published skills contained prompt-injection or unsafe-execution patterns, and the spec so far had no common rules for what an installer should check. Existing skills don&#39;t need any changes.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation&quot;&gt;Read more&lt;/a&gt; — Anthropic&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;harvard%E2%80%93jellyfish-study%3A-ai-agents-lift-code-output-20%E2%80%9330%25-but-resolved-work-only-12%25%2C-as-review-time-rises-49%25&quot; tabindex=&quot;-1&quot;&gt;Harvard–Jellyfish Study: AI Agents Lift Code Output 20–30% but Resolved Work Only 12%, as Review Time Rises 49%&lt;/h2&gt;
&lt;p&gt;A Harvard study using Jellyfish engineering analytics covered &lt;strong&gt;718 firms, about 700,000 workers and roughly 300 million work events&lt;/strong&gt; from January 2021 to March 2026. It pulled data from HR systems, version control, issue trackers and AI tools such as GitHub Copilot, Cursor and Claude Code. It separated two kinds of tools. &lt;strong&gt;AI assistants&lt;/strong&gt; (autocomplete and chat) produced 12% more lines, 9% more commits and 5% more PRs, but only the commit increase was statistically significant. &lt;strong&gt;AI agents&lt;/strong&gt; produced 30% more lines, 20% more commits and 23% more PRs, all significant.&lt;/p&gt;
&lt;p&gt;Completed work tells a different story. With assistants, resolved Jira issues and epics per worker didn&#39;t rise at all. With agents, they rose &lt;strong&gt;12%&lt;/strong&gt;, well below the 30% growth in code. The authors estimate engineers spend only 30–40% of their time writing code. The rest goes to planning, review, testing and deployment, and that is where the extra AI-generated code piled up.&lt;/p&gt;
&lt;p&gt;The review data shows the cost. Average PR review time rose &lt;strong&gt;49%&lt;/strong&gt;, comments per PR rose &lt;strong&gt;53%&lt;/strong&gt;, the share of PRs needing changes &lt;strong&gt;doubled&lt;/strong&gt;, and the share of workers doing reviews rose 14%. Larger PRs didn&#39;t explain the extra effort. The practical lesson for teams adopting agents is that faster code generation doesn&#39;t help much unless review, testing and release keep up. If they don&#39;t, agents move the bottleneck instead of removing it.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/jellyfish-ai-productivity/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;thinkingbox-training%3A-turning-database-state-grading-into-rl-rewards-lifts-pass%401-from-51.7%25-to-60.8%25&quot; tabindex=&quot;-1&quot;&gt;ThinkingBox Training: Turning Database-State Grading Into RL Rewards Lifts pass@1 From 51.7% to 60.8%&lt;/h2&gt;
&lt;p&gt;Microsoft and Hugging Face followed up their ThinkingBox-Bench release with a training recipe. The benchmark grades an agent by checking the &lt;strong&gt;final database state&lt;/strong&gt;, which yields a deterministic pass/fail verdict, and the authors reused those checks as RL rewards on separate training tasks. Before training, they run every candidate task &lt;em&gt;n&lt;/em&gt; times against the base policy, model its success rate with a Jeffreys Beta posterior, and drop tasks it almost never solves (below 0.08) or almost always solves (0.90 and above). In GRPO, groups that all succeed or all fail produce no learning signal, so only the tasks in between are kept.&lt;/p&gt;
&lt;p&gt;They trained &lt;strong&gt;Qwen3.8-27B&lt;/strong&gt; with full-parameter token-mean GRPO for 50 steps on 187 tasks, using 144 episodes per update on 24 H100s. On the held-out 507-task benchmark, &lt;strong&gt;pass@1 rose from 51.70% to 60.78%&lt;/strong&gt;, and the number of tasks passed in all 20 attempts rose from &lt;strong&gt;38 to 107&lt;/strong&gt;. Pass@20, meaning tasks solved at least once, fell slightly, from 453 to 446. The authors call this sharpening: the model became more repeatable, not more capable. One example shows why that matters. On a hotel-booking task that required manager approval, the base model made unauthorized changes in 15 of 20 attempts, and the trained model kept the state intact in all 20.&lt;/p&gt;
&lt;p&gt;The training stack (data selection, rollouts, rewards, Verl integration and a reference 24-H100 config) is open source at microsoft/thinkingbox-training, along with an OpenEnv environment. The authors&#39; three lessons apply to any agent RL project. Probe the base policy&#39;s difficulty distribution before training. Make sure training and evaluation parse tool calls identically. Report both coverage (pass@k) and repeatability (all-k pass), because an agent that solves a task only sometimes is not reliable.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/microsoft/thinkingbox-training&quot;&gt;Read more&lt;/a&gt; — Hugging Face Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h3 id=&quot;safe-%26-secure-ai-agent-practices&quot; tabindex=&quot;-1&quot;&gt;Safe &amp;amp; Secure AI Agent Practices&lt;/h3&gt;
&lt;h2 id=&quot;metr-and-redwood-red-team-openai&#39;s-internal-agent-monitor%3A-coverage-gaps%2C-header-bypasses-and-five-design-ideals&quot; tabindex=&quot;-1&quot;&gt;METR and Redwood Red-Team OpenAI&#39;s Internal Agent Monitor: Coverage Gaps, Header Bypasses and Five Design Ideals&lt;/h2&gt;
&lt;p&gt;METR and Redwood Research published a summary of a three-week pilot (July 21 to August 11) in which they &lt;strong&gt;stress-tested OpenAI&#39;s internal misalignment monitoring system&lt;/strong&gt;. They had access to OpenAI&#39;s codebase, the monitor&#39;s prompts and configuration, and about a dozen engineers. Their estimate is that roughly &lt;strong&gt;75–94%&lt;/strong&gt; of agentic internal Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, and 99.2% for internal Codex agents. Compaction events weren&#39;t monitored, which likely left about 10% of eligible Codex traffic unmonitored from May to July. OpenAI fixed that during the exercise.&lt;/p&gt;
&lt;p&gt;The structural findings apply to anyone building agent monitoring. The monitor ran only when an agent &lt;strong&gt;returned control to the user&lt;/strong&gt;, so an agent that kept acting could postpone review indefinitely. Two client-controllable &lt;strong&gt;headers could disable monitoring&lt;/strong&gt; for new agent instances. Differences between how the agent and the monitor serialized the same input could let an agent craft content that makes the monitor fail without escalating. A redacted exploit could keep hundreds of recent actions out of monitoring. OpenAI says all vulnerabilities affecting production safeguards have been patched; METR hasn&#39;t verified that claim.&lt;/p&gt;
&lt;p&gt;The report sets out five design ideals:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Make monitoring hard to avoid&lt;/strong&gt;: require explicit approval for unmonitored inference, and check agents every N messages or minutes, not only at the end of a turn.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Verify coverage&lt;/strong&gt;: account for all inference, cross-reference it with monitor logs, and make sure the monitor sees exactly the context the agent sees.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Respond quickly&lt;/strong&gt;: clear ownership, defined guidelines and tamper-proof, searchable logs.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Red-team continuously&lt;/strong&gt;: run control evaluations and test for collusion between agents and monitors.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Isolate agents from monitors&lt;/strong&gt;: agent actions must not be able to influence whether or when monitoring runs.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;a href=&quot;https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/&quot;&gt;Read more&lt;/a&gt; — METR&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;metr&#39;s-review-of-anthropic&#39;s-risk-report%3A-low-risk-comes-from-limited-capability%2C-not-proven-safeguards&quot; tabindex=&quot;-1&quot;&gt;METR&#39;s Review of Anthropic&#39;s Risk Report: Low Risk Comes From Limited Capability, Not Proven Safeguards&lt;/h2&gt;
&lt;p&gt;Two days earlier, METR published its review of the &amp;quot;Misalignment in high-stakes settings&amp;quot; section of &lt;strong&gt;Anthropic&#39;s August 2026 Risk Report&lt;/strong&gt;. METR agrees that current misalignment risk is low. But it attributes that mainly to the models&#39; &lt;strong&gt;limited capabilities&lt;/strong&gt;, not to safeguards or alignment training, and it isn&#39;t confident the report&#39;s methods would detect higher risk if it existed.&lt;/p&gt;
&lt;p&gt;The central criticism is about how monitors are evaluated. METR argues that covert-capability benchmarks are a poor proxy for evading a real monitor, because they miss strategies such as &lt;strong&gt;convincing the monitor that a harmful action is legitimate&lt;/strong&gt;. It cites a recent incident in which a model&#39;s chain-of-thought rationalizations reduced how often the monitor flagged it. Some incidents went undetected for months, and Anthropic revised its rating from &amp;quot;very low&amp;quot; to &amp;quot;low&amp;quot; after they came to light. METR also considers some redactions material, especially details about how quickly monitoring responds.&lt;/p&gt;
&lt;p&gt;METR&#39;s recommendations echo its OpenAI findings. Run &lt;strong&gt;end-to-end red-teaming&lt;/strong&gt; of monitoring, for example a well-resourced team trying to start an undetected rogue internal deployment. Add AI systems undermining safety-relevant decision-making to the threat model. Include models still in training in the scope. Publish four specific redacted items. For teams running agents in production, the takeaway is to test monitoring adversarially and end to end rather than assume it works.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://metr.org/blog/2026-10-20-misalignment-risk-report-august-2026-review/&quot;&gt;Read more&lt;/a&gt; — METR&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;docker-sandboxes-%2B-band%3A-multi-agent-teams-with-outbound-only-connectivity-and-proxy-injected-credentials&quot; tabindex=&quot;-1&quot;&gt;Docker Sandboxes + BAND: Multi-Agent Teams With Outbound-Only Connectivity and Proxy-Injected Credentials&lt;/h2&gt;
&lt;p&gt;Docker showed how to connect isolated agents into a team using &lt;strong&gt;BAND&lt;/strong&gt;, a collaboration layer that gives agents persistent identities, shared rooms with humans, @mention routing and visible tool-execution events. Each agent runs in its own &lt;strong&gt;Docker Sandbox microVM&lt;/strong&gt;, with its own kernel, filesystem, Docker daemon and network, and joins a BAND room through the &lt;code&gt;bandhq/band-python-kit&lt;/code&gt; sandbox kit. In the example, a &amp;quot;Repo Doctor&amp;quot; agent fixes a failing test in a mounted workspace when it is mentioned. A hosted planner and research agent share the room but don&#39;t need sandbox access, because only Repo Doctor touches local files.&lt;/p&gt;
&lt;p&gt;The security design is worth copying. &lt;strong&gt;All connectivity is outbound&lt;/strong&gt;: the agent opens a WebSocket to BAND, so nothing has to reach into the sandbox. BAND delivers only messages that mention the agent, which keeps its context narrower than the full room. That also limits how much injected content from other participants it sees. The kit&#39;s network policy allows only BAND and Python package hosts by default. Any other access, such as &lt;code&gt;api.anthropic.com&lt;/code&gt;, has to be granted explicitly with &lt;code&gt;sbx policy allow network&lt;/code&gt;, and &lt;code&gt;sbx policy log&lt;/code&gt; lets you audit actual traffic against the policy.&lt;/p&gt;
&lt;p&gt;Credentials never enter the sandbox. &lt;code&gt;sbx secret set-custom&lt;/code&gt; stores the real BAND and Anthropic keys on the host, binds each one to a single sandbox and destination host, and injects it into outbound requests. Inside the VM the agent sees only a placeholder value. If the agent is compromised, it can&#39;t steal keys it never had, and it can only use them against the hosts it was scoped to.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.docker.com/blog/band-ai-agents-docker-sandboxes/&quot;&gt;Read more&lt;/a&gt; — Docker Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation&quot;&gt;Donating Agent Skills to the Agentic AI Foundation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/jellyfish-ai-productivity/&quot;&gt;Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface.co/blog/microsoft/thinkingbox-training&quot;&gt;Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/&quot;&gt;Red-Teaming OpenAI&#39;s Internal Agent Monitoring Systems&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://metr.org/blog/2026-10-20-misalignment-risk-report-august-2026-review/&quot;&gt;Review of the &amp;quot;Misalignment in high-stakes settings&amp;quot; section in Anthropic&#39;s Risk Report (August 2026)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.docker.com/blog/band-ai-agents-docker-sandboxes/&quot;&gt;Your Agent Runs in a Docker Sandbox. Now It Can Join a Team.&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Anthropic donated the Agent Skills standard (about 1 million published skills, 70+ platforms) to the Linux Foundation&#39;s Agentic AI Foundation, the same home as MCP.</summary>
    
  </entry>
  
  <entry>
    <title>Spring Ecosystem Update: Spock 2.5 Supports Spring Framework 7, Grails 8 Moves Onto Spring Boot Executors, Google ADK Bridges to Spring AI, 2026-10-18</title>
    <link href="https://javarubberduck.com/java/news-2026-10-18-spring/"/>
    <updated>2026-10-18T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-10-18-spring/</id>
    <content type="html">&lt;h2 id=&quot;spock-2.5-adds-spring-framework-7-support-and-grails-8-builds-on-spring-boot&quot; tabindex=&quot;-1&quot;&gt;Spock 2.5 Adds Spring Framework 7 Support and Grails 8 Builds on Spring Boot&lt;/h2&gt;
&lt;p&gt;It was a quiet week for core Spring releases — the Spring blog&#39;s &amp;quot;This Week in Spring&amp;quot; issues for October 6 and 13 carried no new project versions ahead of the next monthly patch cycle — but two Spring-adjacent projects shipped releases that matter for Spring Boot 4 users.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Spock 2.5.0&lt;/strong&gt; adds support for Spring Framework 7 and Groovy 6, which matters for teams that write Spock specifications against Spring Boot 4 applications. The release also makes &lt;code&gt;TooManyInvocationsError&lt;/code&gt; report unsatisfied interactions, so mismatched mock invocations are easier to diagnose, and allows &lt;code&gt;final&lt;/code&gt; local variables in &lt;code&gt;where:&lt;/code&gt; blocks. Breaking changes are listed in the release notes and should be reviewed before upgrading.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Apache Grails 8.0.0&lt;/strong&gt; — which is built on Spring Boot — reached GA with JDK 21 and Gradle 9.7 as minimum versions. All controllers are now REST resources that resolve the controller from the request rather than from the URL mapping, and Grails Promises now run on the Spring Boot task executor, so async work in Grails apps is handled by the same executor infrastructure as the rest of the Boot application.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct05-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;google&#39;s-agent-development-kit-for-java-and-kotlin-ships-spring-ai-bridges&quot; tabindex=&quot;-1&quot;&gt;Google&#39;s Agent Development Kit for Java and Kotlin Ships Spring AI Bridges&lt;/h2&gt;
&lt;p&gt;Google published an overview of its Agent Development Kit (ADK) across five languages — Python, Go, TypeScript, Java and Kotlin — that is directly relevant to Spring shops. The Java edition (&lt;code&gt;adk-java&lt;/code&gt;) targets Maven, Gradle and Spring Boot teams: tools are declared with &lt;code&gt;@Schema&lt;/code&gt; annotations on method parameters, agents are built with a fluent &lt;code&gt;LlmAgent.builder()&lt;/code&gt; API, events stream reactively through RxJava, and a Maven Dev UI plugin is included. Crucially, ADK for Java ships &lt;strong&gt;bridges for Spring AI and LangChain4j&lt;/strong&gt;, so ADK agents can reuse models and tools already wired up in an existing Spring AI application.&lt;/p&gt;
&lt;p&gt;The new Kotlin edition (&lt;code&gt;adk-kotlin&lt;/code&gt;) uses a KSP compiler plugin to read &lt;code&gt;@Tool&lt;/code&gt; and &lt;code&gt;@Param&lt;/code&gt; annotations at build time and emit tool schemas into bytecode, so unsupported parameter types fail the Gradle build rather than at runtime. It also has JVM server support with a Spring AI bridge and experimental ADK 2.0 graph workflows.&lt;/p&gt;
&lt;p&gt;All five languages share an MCP client and support the Agent2Agent (A2A) protocol, which means a Spring Boot service written with ADK for Java can call tools exposed by — or delegate to — agents written in Python or Go. Google&#39;s &lt;code&gt;agents-cli&lt;/code&gt; scaffolder can generate projects and deploy them to Cloud Run or GKE.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://cloud.google.com/blog/topics/developers-practitioners/one-agent-development-kit-five-languages&quot;&gt;Read more&lt;/a&gt; — Google Cloud Blog&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct05-2026/&quot;&gt;Java News Roundup: Grails 8, Open Liberty, GlassFish, Infinispan, LangChain4j, Spock&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://cloud.google.com/blog/topics/developers-practitioners/one-agent-development-kit-five-languages&quot;&gt;One Agent Development Kit, five languages&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Spock 2.5.0 adds Spring Framework 7 and Groovy 6 support, while Apache Grails 8.0.0 runs on JDK 21 and moves Grails Promises onto the Spring Boot task executor.</summary>
    
  </entry>
  
  <entry>
    <title>JavaRubberDuck Podcast — October 18, 2026</title>
    <link href="https://javarubberduck.com/podcast/2026-10-18/"/>
    <updated>2026-10-18T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/podcast/2026-10-18/</id>
    <content type="html">&lt;h2 id=&quot;in-this-episode&quot; tabindex=&quot;-1&quot;&gt;In This Episode&lt;/h2&gt;
&lt;p&gt;Stories covered from 2026-10-18 to 2026-10-18:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Gemini 4 Argon Is Generally Available on Google Cloud and in GitHub Copilot&lt;/strong&gt; — Google Cloud Blog (&lt;a href=&quot;https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Claude Code 2.1.299–2.1.302: Idle Compaction, Fork Subagents Everywhere, Tighter Hook Validation&lt;/strong&gt; — Claude Code Docs (&lt;a href=&quot;https://code.claude.com/docs/en/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Codex CLI 0.163 Adds a Searchable Config Panel and Mid-Response Steering; GPT-5.5 Retires From Codex&lt;/strong&gt; — OpenAI (&lt;a href=&quot;https://learn.chatgpt.com/docs/changelog&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Copilot: HydraFusion Max and Agentic Autofix for Code Scanning Reach Users&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-14-hydrafusion-updates-and-new-max-option&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Reflection Releases Beam, a 501B Apache-2.0 Open Model for Agentic Coding&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/reflection-beam&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Meta Publishes Muse Spark Open, an 816B Multimodal Open-Weight Model&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/meta-muse-spark-open&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hermes Agent v0.22.0 &amp;quot;The Threshold Release&amp;quot; Adds Guided Setup, App Connectors and Local Models&lt;/strong&gt; — GitHub (&lt;a href=&quot;https://github.com/NousResearch/hermes-agent/releases&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Four Eval Engineering Lessons From Building Agent Plugins: Grade the Trajectory, Not the Summary&lt;/strong&gt; — Google Cloud Blog (&lt;a href=&quot;https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Grafana Applies SLOs and Error Budgets to AI Agent Behaviour&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/grafana-slo-ai-agent/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;IBM Granite Switch 4.1 Serves Chat, Decisions and Speech From One Endpoint&lt;/strong&gt; — Hugging Face (&lt;a href=&quot;https://huggingface.co/blog/ibm-granite/granite-switch-4-1&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Stack Overflow 2026 Survey: Only 7% of Developers Trust AI Output for Important Decisions&lt;/strong&gt; — Docker (&lt;a href=&quot;https://www.docker.com/blog/stack-overflow-survey-2026-ai-agent-trust/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Chainguard Launches Workspaces and Checks: microVM Sandboxes and Attested CI for Coding Agents&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/chainguard-workspaces-checks/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;AWS Strands Box: Kernel-Enforced Agent Isolation With Dogwood Policies Outside the Model&lt;/strong&gt; — InfoQ (&lt;a href=&quot;https://www.infoq.com/news/2026/10/aws-strands-agents/&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Anthropic Gives METR Employee-Like Access and Publishes a Three-Tier Model for Independent Evaluators&lt;/strong&gt; — Anthropic (&lt;a href=&quot;https://www.anthropic.com/news/an-update-on-independent-evaluators&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;GitHub Actions Workflow Lockfiles Pin Actions and Reusable Workflows to Exact Commits&lt;/strong&gt; — GitHub Changelog (&lt;a href=&quot;https://github.blog/changelog/2026-10-14-github-actions-workflow-lockfiles-in-public-preview&quot;&gt;link&lt;/a&gt;)&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;listen&quot; tabindex=&quot;-1&quot;&gt;Listen&lt;/h2&gt;
&lt;audio controls=&quot;&quot;&gt;
  &lt;source src=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-10-18-podcast.mp3&quot; type=&quot;audio/mpeg&quot; /&gt;
  &lt;a href=&quot;https://pub-4fe17297251e45d1974e07eb5ac68578.r2.dev/2026-10-18-podcast.mp3&quot;&gt;Download episode&lt;/a&gt;
&lt;/audio&gt;
&lt;blockquote&gt;
&lt;p&gt;🤖 &lt;strong&gt;AI-generated episode.&lt;/strong&gt; Both hosts are synthetic voices produced with Google NotebookLM — no human narrators were recorded. The conversation is generated from the news digest articles linked above.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;subscribe&quot; tabindex=&quot;-1&quot;&gt;Subscribe&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://javarubberduck.com/podcast.xml&quot;&gt;RSS feed&lt;/a&gt; — paste into any podcast app&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://open.spotify.com/show/&quot;&gt;Spotify&lt;/a&gt; — coming soon&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>Gemini 4 Argon goes GA, Reflection&#39;s 501B Beam and Meta&#39;s Muse Spark Open land as open weights, and new sandboxes, SLOs and lockfiles for keeping AI coding agents in check.</summary>
    
  </entry>
  
  <entry>
    <title>Java News: Generational Shenandoah Becomes the Default, JDK 28 Locks In Structured Concurrency and ZGC Upgrades, Helidon 27 Goes Tip and Tail, 2026-10-18</title>
    <link href="https://javarubberduck.com/java/news-2026-10-18-java-jdk/"/>
    <updated>2026-10-18T00:00:00.000Z</updated>
    <id>https://javarubberduck.com/java/news-2026-10-18-java-jdk/</id>
    <content type="html">&lt;h2 id=&quot;jep-535-targeted%3A-generational-shenandoah-becomes-the-default-in-jdk-28&quot; tabindex=&quot;-1&quot;&gt;JEP 535 Targeted: Generational Shenandoah Becomes the Default in JDK 28&lt;/h2&gt;
&lt;p&gt;JEP 535, &amp;quot;Shenandoah GC: Generational Mode by Default,&amp;quot; has moved to &lt;strong&gt;Targeted&lt;/strong&gt; for JDK 28. The change is small in code but significant in behaviour: the default value of &lt;code&gt;ShenandoahGCMode&lt;/code&gt; switches from &lt;code&gt;satb&lt;/code&gt; to &lt;code&gt;generational&lt;/code&gt;. From JDK 28, plain &lt;code&gt;-XX:+UseShenandoahGC&lt;/code&gt; runs Generational Shenandoah, while explicitly selecting &lt;code&gt;-XX:ShenandoahGCMode=satb&lt;/code&gt; still runs the classic single-generation collector but prints a deprecation warning. The non-generational mode is not removed yet — the JEP only deprecates it with the intent to remove it in a later release.&lt;/p&gt;
&lt;p&gt;The motivation is twofold. Generational mode collects young objects frequently and old objects rarely, which the JEP says gives lower concurrent overhead, better allocation throughput and a smaller memory footprint for most workloads. And maintaining two Shenandoah modes in parallel has been slowing down new feature work in the collector.&lt;/p&gt;
&lt;p&gt;For teams running Shenandoah in production, the practical work is mostly in tooling: GC log output and the data exposed via serviceability and management APIs may differ, and anything that parses JVM output could trip over the new warning. Teams that depend on the old mode for a specific latency profile should pin &lt;code&gt;-XX:ShenandoahGCMode=satb&lt;/code&gt; explicitly now and start benchmarking the generational mode on JDK 28 early-access builds.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://openjdk.org/jeps/535&quot;&gt;Read more&lt;/a&gt; — OpenJDK&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;jdk-28-feature-set-grows-to-10-targeted-jeps%2C-including-final-structured-concurrency&quot; tabindex=&quot;-1&quot;&gt;JDK 28 Feature Set Grows to 10 Targeted JEPs, Including Final Structured Concurrency&lt;/h2&gt;
&lt;p&gt;The JDK 28 project page now lists ten JEPs as targeted. Beyond JEP 535, the most notable status change since last week is that the three JEPs that were Proposed to Target in early October — &lt;strong&gt;JEP 543 (Structured Concurrency)&lt;/strong&gt;, &lt;strong&gt;JEP 545 (Faster Startup and Warmup with ZGC)&lt;/strong&gt; and &lt;strong&gt;JEP 546 (Adaptive Heap Sizing for ZGC)&lt;/strong&gt; — all completed their review periods and are now formally &lt;strong&gt;Targeted&lt;/strong&gt;. JEP 543 finalizes &lt;code&gt;StructuredTaskScope&lt;/code&gt; without change after two incubator rounds (JDK 19–20) and seven previews (JDK 21–27), so code written against the JDK 27 preview API should compile on JDK 28 without &lt;code&gt;--enable-preview&lt;/code&gt;.&lt;/p&gt;
&lt;p&gt;The full targeted list is now: JEP 401 Value Objects (Preview), JEP 535 Generational Shenandoah by default, JEP 539 Strict Field Initialization in the JVM (Preview), JEP 540 Simple JSON API (Incubator), JEP 541 Deprecate the macOS/x64 Port for Removal, JEP 542 PEM Encodings of Cryptographic Objects, JEP 543 Structured Concurrency, JEP 544 Ahead-of-Time Code Compilation, JEP 545 and JEP 546.&lt;/p&gt;
&lt;p&gt;Waiting in the wings, &lt;strong&gt;JEP 547, Lazy Constants (Fourth Preview)&lt;/strong&gt;, moved from draft to Candidate. The API — formerly known as Stable Values and Computed Constants — provides immutable values that are initialized at most once, and this round proposes a fourth preview with no API changes. JDK 28 early-access Build 19 is also available for testing.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://openjdk.org/projects/jdk/28/&quot;&gt;Read more&lt;/a&gt; — OpenJDK&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;helidon-27-adopts-oracle&#39;s-tip-and-tail-model-with-new-messaging-and-data-modules&quot; tabindex=&quot;-1&quot;&gt;Helidon 27 Adopts Oracle&#39;s Tip and Tail Model With New Messaging and Data Modules&lt;/h2&gt;
&lt;p&gt;Helidon 27 is the first release of Oracle&#39;s microservices framework to follow the &amp;quot;Tip &amp;amp; Tail&amp;quot; model of library development. Helidon 27 is the &lt;strong&gt;tip&lt;/strong&gt;, aligned with JDK 27 and released on the OpenJDK cadence (Helidon 28 is expected alongside JDK 28 in March 2027). Helidon 4.5.5 is the &lt;strong&gt;tail&lt;/strong&gt;: it stays on the 4.5 LTS line with a JDK 21 baseline, adds JDK 25 support, and receives only critical fixes, security patches and selected backports.&lt;/p&gt;
&lt;p&gt;The programming models were restructured. &lt;strong&gt;Helidon Core&lt;/strong&gt; replaces Helidon SE and offers two styles: an imperative model in plain Java with no injection or annotations, and a declarative model using dependency injection and annotations whose wiring is generated at build time. &lt;strong&gt;Helidon MicroProfile&lt;/strong&gt; replaces Helidon MP and moves into its own project, while the former SE Integrations become &lt;strong&gt;Helidon Extensions&lt;/strong&gt; with independent release lifecycles.&lt;/p&gt;
&lt;p&gt;Two new modules stand out. &lt;strong&gt;Helidon Messaging&lt;/strong&gt; offers a blocking (rather than reactive) messaging layer — natural on virtual threads — with Kafka, JMS and Apache Pulsar connectors bound through configuration. &lt;strong&gt;Helidon Data JDBC&lt;/strong&gt; provides a repository-query API whose query interfaces are turned into implementation classes at compile time, and supports Jakarta Persistence via EclipseLink and Hibernate ORM.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/helidon-27-released/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;java-ecosystem-roundup%3A-open-liberty-on-jdk-27%2C-langchain4j-1.22%2C-micronaut-netty-cve-fixes&quot; tabindex=&quot;-1&quot;&gt;Java Ecosystem Roundup: Open Liberty on JDK 27, LangChain4j 1.22, Micronaut Netty CVE Fixes&lt;/h2&gt;
&lt;p&gt;InfoQ&#39;s roundup for the week of October 5 collected a busy set of ecosystem releases. &lt;strong&gt;Open Liberty 26.0.0.10&lt;/strong&gt; reached GA with JDK 27 support and fixes for seven CVEs; because JDK 27 makes G1 the default collector in all environments, the release notes warn that containers with constrained resources may see different GC behaviour than on earlier Java versions.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;LangChain4j 1.22.0&lt;/strong&gt; added a &lt;code&gt;BedrockBatchChatModel&lt;/code&gt; that implements &lt;code&gt;BatchChatModel&lt;/code&gt; on top of the Amazon Bedrock batch inference API, plus &lt;code&gt;OpenAiDecisionModel&lt;/code&gt; and &lt;code&gt;OpenAiOfficialDecisionModel&lt;/code&gt; for OpenAI&#39;s new Decisions API. &lt;strong&gt;Micronaut 5.2.2&lt;/strong&gt; upgrades Netty to 4.2.19.Final, resolving 16 CVEs including HTTP request/response smuggling and unbounded resource consumption in the HTTP/1, HTTP/2 and HTTP/3 codecs — a patch worth applying promptly.&lt;/p&gt;
&lt;p&gt;Other releases: &lt;strong&gt;Infinispan 16.3.0&lt;/strong&gt; improves hash-based conflict resolution when merging partitions and persists configuration attributes across restarts; &lt;strong&gt;GlassFish 8.0.5&lt;/strong&gt; speeds up deployments and fixes ZIP64 extra-field headers; and &lt;strong&gt;TornadoVM 7.2.0&lt;/strong&gt; adds &lt;code&gt;mmaLoadA()&lt;/code&gt;/&lt;code&gt;mmaLoadB()&lt;/code&gt; matrix-multiply fragment loaders for &lt;code&gt;HalfFloat&lt;/code&gt; and more accurate CUDA dispatch timing.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct05-2026/&quot;&gt;Read more&lt;/a&gt; — InfoQ&lt;/p&gt;
&lt;hr /&gt;
&lt;h2 id=&quot;links-%26-sources&quot; tabindex=&quot;-1&quot;&gt;Links &amp;amp; Sources&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://openjdk.org/jeps/535&quot;&gt;JEP 535: Shenandoah GC: Generational Mode by Default&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://openjdk.org/projects/jdk/28/&quot;&gt;JDK 28 project page&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/helidon-27-released/&quot;&gt;Helidon 27 Introduces Restructured Programming Models and New Messaging and Data Modules&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.infoq.com/news/2026/10/java-news-roundup-oct05-2026/&quot;&gt;Java News Roundup: Grails 8, Open Liberty, GlassFish, Infinispan, LangChain4j, Spock&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
</content>
    
    <summary>JEP 535 is now targeted to JDK 28, making Generational Shenandoah the default Shenandoah mode and deprecating the non-generational mode.</summary>
    
  </entry>
  
</feed>
