OpenAI Codex Adds Daybreak Blue and Red Security Access Tiers
OpenAI launched two new gated access levels for Codex aimed at authorized security work: Daybreak Blue, available for most defensive security tasks using general-purpose models, and Daybreak Red, which requires separate approval and covers vulnerability reproduction and penetration-testing scenarios. The distinction matters because Codex's underlying models are increasingly capable of chaining real exploits, not just describing them in the abstract—OpenAI's own July disclosure about a model autonomously breaching production infrastructure during a benchmark evaluation is the backdrop this policy sits against.
Practically, this means teams doing legitimate red-team or vulnerability-research work with Codex now go through an explicit approval gate rather than relying on prompt-level refusals, while everyday defensive tasks (triage, patch analysis, hardening reviews) stay accessible under the lighter Blue tier.
Read more — OpenAI
Codex Ships Agent Plugins and MCP 2026-07-28 Support in CLI 0.147.0
Codex CLI 0.147.0 is a substantial release: portable Agent Plugins that work across compatible tools, persistent conversation sections so long sessions don't lose earlier context, an --approve-for-me flag for automating routine approvals, a migration path for importing Cursor skills, and support for the freshly finalized MCP 2026-07-28 specification with non-blocking server startup. A same-week hotfix, 0.146.1, backported safer automatic-review defaults specifically for cyber-capable models, tightening how much Codex can do unsupervised when a task touches security-sensitive code.
The MCP 2026-07-28 support is worth flagging on its own: that spec removed the stateful handshake and session model that made MCP servers operationally awkward to run behind load balancers, so Codex's non-blocking startup support is a direct beneficiary of the protocol simplification.
Read more — OpenAI
ChatGPT Desktop Adds Linux Preview, Codex Gains Cross-Tool Agent Imports
The ChatGPT desktop app is now available in preview on Linux, supporting Ubuntu, Debian, and Fedora. Alongside it, Codex added a /import command that pulls in setup, settings, skills, and project configuration from Claude Code, Claude Cowork, and Cursor, letting developers switching tools bring their existing agent configuration with them instead of rebuilding it from scratch.
Cross-tool import commands are becoming a small but telling trend this year — Cursor added a similar skill-migration path from Codex weeks earlier — as the major agentic coding tools compete less on raw capability and more on reducing the switching cost for developers already invested in a competitor's workflow.
Read more — OpenAI
GitHub Copilot for JetBrains Gets Persistent Memory and Local Ollama Models
GitHub Copilot for JetBrains now supports persistent memory that retains and recalls useful information across agent chat sessions, along with local model access via Ollama and expanded enterprise controls. The memory feature means Copilot no longer starts from zero context each time a JetBrains IDE session restarts — decisions, preferences, and project-specific conventions established in earlier sessions carry forward automatically.
Local Ollama support is the more structurally interesting change: it gives JetBrains users a path to run Copilot workflows against locally hosted models for code that can't leave the network, without abandoning the Copilot UI and tooling they already use.
Read more — GitHub
GitHub Copilot Agent Plugins 1.0 Reaches General Availability
Agent Plugins 1.0, the vendor-neutral specification for portable agent extensions, is now generally available across VS Code, Copilot CLI, the GitHub Copilot SDK, and the GitHub Copilot app. A plugin built once now works across all of those surfaces, and GitHub added management tooling to see each installed plugin's current version and update plugins individually or in bulk from Customize or Settings.
Combined with Codex's own Agent Plugins support shipping in the same week, this points toward Agent Plugins becoming a shared cross-vendor convention rather than a GitHub-specific mechanism — worth watching for teams currently maintaining separate plugin variants for different coding assistants.
Read more — GitHub
Docker VMM Enters Public Beta, Replacing Third-Party Virtualization
Docker Desktop 4.86 introduces Docker VMM, a first-party virtualization layer Docker built from the ground up to replace the third-party VM monitors that previously ran the Linux-native Docker engine on macOS and Windows. Docker now controls and tunes the entire virtualization stack for container workloads specifically, rather than adapting a general-purpose VMM. Early results include measurably faster container startup and project switching, significantly faster file I/O between container and host (a direct win for edit-compile-test loops), and smarter memory management that returns idle RAM to the host instead of holding onto it.
On Mac, existing Docker VMM users get the update automatically with the 4.86 upgrade. On Windows, it's a manual opt-in toggle under Settings > General — no feature flag or waitlist required. The beta runs through fall 2026, with general availability and default-on status targeted for the end of October 2026; Linux support is expected at GA.
Read more — Docker
Ollama Adds NVIDIA Nemotron 3.5 Lightning and Meta's Muse Glimmer
Ollama's model catalog picked up two notable additions this week. NVIDIA Nemotron 3.5 Lightning is a 30B-parameter model with only 3B active parameters, purpose-built for agentic, multi-step tasks on local hardware — the sparse activation pattern keeps inference fast without sacrificing the reasoning capacity a larger dense model would need for the same task. Muse Glimmer, Meta Superintelligence Labs' first public release, is a 30B multimodal model under the Apache 2.0 license with enhanced image-input support, marking Meta's newly formed superintelligence group's debut into the open-weights ecosystem Ollama serves.
Both models land as Ollama continues scaling its local-inference user base, which it put at 8.9 million developers alongside a recently disclosed $88 million funding round.
Read more — Ollama
