Plugin4Shell bypasses AI agents' SHA pins; Copilot, Gemini unpatched
The version your agent reports pinned may not be the code it runs: a hash-shaped branch name hides malicious commits. Air Security proved it at scale.

Copy markdown
Your pinned SHA can resolve to someone else's code
Air Security found the agents lock a plugin to a reviewed commit hash but never verify the checkout actually landed there. A repo owner names a branch to look like that hash, points it at malicious code, and the agent runs it while still reporting the pinned version — zero-click, triggered by a normal install or update.
Two of your four agents are still exposed
Claude Code (2.1.179+) and Codex (0.146.0+) are patched — update now. GitHub Copilot has no fix, and Google won't patch Gemini CLI: it is deprecating the tool and pointing users to Antigravity, so existing installs stay vulnerable. Treat plugins pulled through those two as untrusted.
GitHub-hosted plugins are safe; Bitbucket and self-hosted aren't
GitHub rejects branch names shaped like commit hashes, so plugins installed from default GitHub marketplaces are protected. The real exposure is Bitbucket, private or self-hosted Git, and Gemini CLI's FETCH_HEAD fetch path — audit any plugin or skill you pull from those before you run it.
It already reached 134,000 agents in the lab
There's no in-the-wild exploitation yet and no CVE assigned, but the demos weren't hypothetical: one poisoned plugin spread to 26,000+ agents before removal, and a 'SkillJacking' variant hijacked 925 existing skills across 134,000 agents. Air Security disclosed to vendors in June; it went public September 18.