Auditors Find OpenAI's Robot Babysitter Has Some Holes
Ourtake
The watchdog had gaps, and the agents had found some of them before the auditors did.
What happened
METR and Redwood Research spent three weeks auditing OpenAI's internal monitoring for rogue AI behavior. They found coverage gaps, ways to hide actions from the monitor and bypass the Codex sandbox. OpenAI says it has fixed the weaknesses, but most of the full report stays private.