IntelliJ IDEA 2026.2 Ships with Day-One Java 27 and Kotlin 2.4 Support
JetBrains has released IntelliJ IDEA 2026.2, timed to support Java 27's latest language features and Kotlin 2.4 from day one — useful for teams tracking the JDK closely, since Java 27 itself isn't due until September. The release also adds TypeScript 7.0 compatibility and Terraform testing framework support.
Debugging gets two concrete upgrades: logpoints, which let developers insert breakpoint-style log statements without modifying source code, and enhanced navigation that lets you jump directly from runtime output back to the originating source line. Spring developers get streamlined database migration workflows and richer Spring Security insight panels, continuing JetBrains' pattern of pairing every major release with Spring-specific tooling improvements.
On the AI side, IntelliJ IDEA 2026.2 adds native GitHub Copilot integration, an agent skills capability (compatible with the same Agent Skills format gaining traction across the industry), and AI code completion support for third-party model providers beyond JetBrains' own. The release also includes more than 1,300 bug fixes and usability improvements, along with fixes for roughly 140 freezes and performance issues — a substantial stability pass alongside the feature work.
Read more — JetBrains Blog
Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open Model
Moonshot AI has released Kimi K3, an open-weight Mixture-of-Experts transformer with 2.8 trillion total parameters and roughly 50 billion active per token, routing through 16 of 896 experts. The model supports a 1-million-token context window and native vision input alongside text.
K3 introduces three architectural changes worth noting for anyone building on top of it. Kimi Delta Attention (KDA) is a hybrid linear-attention mechanism that keeps full expressiveness in the layers that need it while cutting computational cost elsewhere — important for making a 1M-token context window tractable. Attention Residuals (AttnRes) let layers pull representations from arbitrary earlier layers rather than only the immediately preceding one, which the team reports helps particularly in MoE architectures. Stable LatentMoE handles expert routing in latent space with quantile balancing and soft dropout to manage overflow, aimed at keeping training stable at this scale.
Rather than quantizing after training, K3 was trained with quantization-aware training from the start: weights ship in MXFP4 (4-bit floating point with per-block scaling), cutting storage to around 1.4TB — about a quarter of the FP16 footprint — while activations use MXFP8. Both formats have native hardware support on NVIDIA Blackwell and AMD MI400 accelerators, so the efficiency gains aren't just theoretical. On benchmarks, K3 leads on SWE Marathon and Program Bench for coding, scores 1,687 on GDPval-AA v2, and hits 95.0 F1 on DeepSearchQA for agentic tasks. The open-weight release makes a model at this parameter count deployable for mid-sized organizations for the first time, though the writeup flags known weaknesses around sensitivity to prior "thinking" history and over-eager behavior in ambiguous prompts.
Read more — Hugging Face
Docker Publishes a Practical Guide to Building AI Agents Safely
Docker's developer relations team published a detailed explainer on what AI agents actually are and how to build with them without creating a security incident. The framing is direct: agents differ from chatbots because they operate in a loop — perceive, reason, act, observe, adapt — driven by three properties: autonomy (deciding the next action without step-by-step approval), tool use (executing code, calling APIs, changing files), and memory (carrying context across steps).
The core argument is that safety is an infrastructure problem, not a model problem — the reasoning model cannot contain itself, so containment has to happen at the environment layer. Docker's recommendation is to treat autonomy as a dial rather than a switch: the more autonomy you grant an agent, the larger its potential blast radius, so tool access should be scoped to exactly what a given task needs, and agents should run in isolated, disposable environments so a bad decision can't reach host credentials, other projects, or real files. The post specifically calls out microVM-based isolation as a way to get strong isolation without a meaningful performance penalty, and references the now-familiar "rm -rf incident" pattern — an agent without containment deleting the wrong directory — as the failure mode this architecture is meant to prevent.
On the tooling side, the post points to Docker Sandboxes (isolated, disposable microVMs per agent, letting an agent run in "YOLO mode" safely), Docker AI Governance (centrally defined rules for allowed actions, network reach, credentials, and tools, enforced across a team), Docker Model Runner for local-first inference, and the Docker MCP Catalog and Toolkit for managing MCP tool connections. The closing line captures the thesis well: the hard part isn't getting an agent to do something, it's deciding how much it's allowed to do and where.
Read more — Docker Blog
OpenAI Codex Ships Command-Safety and Desktop Integration Updates
OpenAI's Codex CLI continues shipping frequent point releases. Version 0.144.5, released July 16, improves dangerous-command detection, specifically catching more forced variants of rm before they execute — a direct response to the kind of destructive-agent incidents that have prompted safety guidance across the industry this year. Earlier in the same release train, 0.144.1 and 0.144.0 added usage-limit credits, new app-approval modes, and improved MCP tool authentication.
Codex also landed in the ChatGPT desktop app (build 26.707) with direct Markdown editing and inline annotations, a PR Chat feature for reviewing GitHub pull requests without leaving the app, custom domain support for published Sites, and performance improvements from GPT-5.6 on Computer Use tasks. On mobile, inline visualizations were added to Codex tasks, along with more reliable linking when a task is created from within a conversation.
Read more — OpenAI Developers
Links & Sources
You May Also Enjoy
