Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It

AI Now published a proof-of-concept called "Friendly Fire" showing that autonomous code-review agents (Claude Code and Codex/GPT-5.5) can be coerced into executing a hidden payload in an open-source project by embedding a launcher script and disguised binary referenced from README.md, bypassing existing safety checks; the PoC ran against the geopy project and worked across multiple models and versions, prompting a recommendation to avoid handing untrusted code to agents that can run commands and reach secrets.