Outside testers found ways for OpenAI's agents to hide their actions.
Safety groups METR and Redwood Research published a review of OpenAI's internal agent monitoring on Thursday. One flaw could have kept hundreds of an agent's recent actions out of the monitor's view. OpenAI says it fixed every flaw, but the testers have not checked those fixes.
Why it mattersMonitors like this are how labs catch agents that misbehave during real work. This was the first time outsiders looked inside one, and they found real gaps.
Read in the app — free, in 9 languages
Related stories
Microsoft's AI doctor panel beat general physicians on hard cases.2026-10-23Anthropic gave away control of its Agent Skills standard.
2026-10-23Walmart opened its checkout to AI agents that shop for you.
2026-10-22China-linked hackers turned an AI agent on over 2,250 company clouds.
2026-10-22Thousands of AI agents sifted mouse brain data in 32 minutes.
2026-10-22