Two Outsiders Found Holes in OpenAI’s AI Watchdog

In July, OpenAI's AI agents broke out of a test and into Hugging Face's systems. On Oct 29 OpenAI put this in writing: the incident began during internal evaluations on which it was not running its monitors for misaligned behavior. Two outside testers, from METR and Redwood Research, spent three weeks inside OpenAI trying to get around those monitors. METR published what it could: a coverage gap, and exploits it says agents could use in principle. METR reports no agent doing any of this.