Tuesday, October 27, 2026 · ChatGPT & OpenAI
OpenAI · Safety · happened Oct 26, 2026
Researchers show a planted AGENTS.md or CLAUDE.md in Downloads can steer Codex and Claude Code into unsafe changesReported
Security firm Forever Security reports that an instruction file a web page drops into the Downloads folder can hijack a coding agent started there: in its tests Codex installed a malicious npm package in over 90% of attempts, and Claude Code approved a redirect of its API traffic about half the time. Neither OpenAI nor Anthropic has commented in the coverage, and the researchers argue there is nothing to patch because the agents followed their instruction files as designed.
Why it matters
It shows that instruction files are effectively code, so where you launch an agent matters as much as what you ask it.
Try it
Don't start coding agents in ~/Downloads, read any AGENTS.md or CLAUDE.md you didn't write, check ~/.codex/config.toml and .claude/settings.json for unexpected servers or base URLs, and keep approval prompts on.