Outside evaluators found ways AI agents could have bypassed OpenAI’s earlier monitoring. METR and Redwood Research published findings on October 22 from a three-week summer review.
One reported attack path could have hidden hundreds of recent actions from monitoring. Another concerned sandboxing and automatic approval checks. The findings do not establish that agents actually used those opportunities.
The two evaluators assessed the system as it existed on and before July 21. OpenAI subsequently said all identified weaknesses affecting production safeguards had been patched. METR and Redwood have not verified that statement.
OpenAI received a 61-page internal report. Only a summary is public, with the company allowed to redact material. The exercise therefore measures neither today’s complete coverage nor the effectiveness of all subsequent fixes.