OUTSIDE AUDITORS SPENT 3 WEEKS INSIDE OPENAI'S AI WATCH SYSTEM — FOUND THE OFF SWITCHES

Two outside researchers — one from METR, one from Redwood Research — spent three weeks inside OpenAI with deep access to its codebase and monitoring configuration, trying to break the system that watches the company's own AI agents for scheming and sabotage. Their 61-page report found the watchers had blind spots.

Reinforcement-learning training runs were not monitored at the start of the exercise. Checks only fired when an agent handed control back to the user, so an agent could delay review indefinitely by simply continuing to act. Two client-controllable settings could switch monitoring off for new agents. And the pair found an exploit — details redacted by OpenAI — that could keep hundreds of recent agent actions out of the monitor's sight.

OpenAI says every flaw affecting its production safeguards has been patched. METR has not verified that claim.