Two outside researchers spent three weeks inside OpenAI this summer, hunting for ways its AI agents could slip past the system built to watch them. Their full report ran 61 pages.
OpenAI’s researchers and security staff got all of it. The auditors’ own bosses at METR and Redwood Research, two AI safety groups, got a 38-page version OpenAI had redacted.
The public got a short summary. OpenAI could cut anything it considered sensitive, and the auditors couldn’t describe the cuts on their own.
- An agent could hide hundreds of its recent actions from the monitor. How it works is redacted by OpenAI.
- Two settings could switch off monitoring for any new agent, though flipping them might get caught.
- Roughly 10% of internal work by OpenAI’s coding agents went unwatched from May to July. OpenAI had already spotted it and fixed it during the audit.
OpenAI says it has since patched every hole that touched its real safeguards, which METR hasn’t checked.
A few weeks ago, OpenAI fired three safety researchers, including its point person for METR’s investigation into the Hugging Face hack.
OpenAI says they mishandled sensitive information. They say they followed the rules and warned the firings could become an excuse to give auditors less access.
