Vulnerability / research

METR and Redwood find gaps in historical OpenAI agent safeguards

Reported

Published here

What happened

A high-access research exercise identified ways internal agents could bypass sandbox controls or evade monitoring, including hiding recent actions. The findings concern OpenAI’s system as it existed on or before July 21.

Parties and reporting

AI tools and systems: Internal Codex agents, OpenAI Responses API, Internal misalignment monitoring

Outcome

OpenAI says all identified vulnerabilities affecting production safeguards were patched. METR and Redwood have not evaluated that claim or the expanded monitoring coverage.

Operating lesson

Isolate monitors from agent-controlled systems, reconcile monitoring logs against inference traffic, and test coverage during long-running tasks.

Evidence and limits

Published October 22; testing ran July 21–August 11. This controlled review is separate from the October 20 alert-suppression incident and the Hugging Face investigation. Some proposed paths were not validated; the exercise did not establish real-world exploitation or current system coverage.

Sources