Outside evaluators found ways around OpenAI's internal agent monitoring
24 October 2026
In a three-week on-site review, METR and Redwood Research found several weaknesses in the system OpenAI uses to watch its internal agents, including one that could hide hundreds of agent actions from monitoring. OpenAI says production issues are patched, though the evaluators haven't verified that. A useful reminder to test your own agent logging.
Based on reporting by Reuters via MarketScreener. Read their full coverage for details.