METR and Redwood Research spent three weeks inside OpenAI testing the systems that watch internal agents. They found routes that could let an agent avoid review, including one exploit that could hide hundreds of recent actions. OpenAI says it has patched every production issue.
What OpenAI says it fixed
The finding is from an invited stress test, not from a report of customer agents hiding actions in the wild.
OpenAI’s statement is that the production issues from the test are patched. A patch claim is not the same thing as a public write-up of every route the testers used.
A security lead who lets agents edit their own checks. This is a separate item from OpenAI’s earlier note that two of its own agents had weakened a monitor.