Air Security disclosed Plugin4Shell (Sept 17): agents check out a marketplace-pinned plugin SHA but never verify HEAD, so a malicious branch named like the pin can swap in code on auto-update—zero-click RCE. Claude Code (≥2.1.179) and Codex (≥0.146.0) are patched; Copilot has no fix yet, and deprecated Gemini CLI will not be patched.
The UN launched System Data Commons (built on Google’s open source Data Commons) with MCP so agents can pull authoritative stats from 20 agencies at launch, with lineage back to the source. A UNICEF benchmark found six frontier models averaged just 21.2% accuracy on development indicators without that grounding…
GitLab 19.4 adds beta MCP tools so agents can run pipelines, open/update merge requests, triage vulnerabilities, and manage work items—with read only actions auto allowed and writes gated by the same tool governance as Duo. A new “merge request created” trigger can fire flows as soon as a diff is ready…
GitHub ported the Copilot agent runtime from TypeScript/Node to 832K lines of production Rust across 128 PRs, with agents writing most of the code. One developer led the 14.5 week in place migration; in process one turn client startup fell from 5.25s to 292ms…
Anthropic is merging Cowork into the main Claude app so Claude picks chat vs longer agentic work, and shipping beta Docs and Slides editors for paid plans. Pro and Max get the rollout first over the next few weeks; Claude Code stays separate…
Anthropic opened a Pro/Max beta that turns Projects into a coordinator: it splits an engineering goal across parallel Claude Code cloud sessions (each on its own branch), shares project memory, and reviews/assembles results. Each thread counts as a full session, so parallel work burns plan usage faster…
Spotted something we missed? Start a thread.
