Red-Teaming OpenAI's Internal Agent Monitoring Systems

What happened: METR and Redwood Research ran a three-week pilot exercise, from July 21st to August 11th, stress-testing OpenAI's internal misalignment monitoring system and produced a 61-page internal report. Key numbers: 61-page report; roughly 75 to 94% of agentic internal Responses API Why it matters: The exercise found novel vulnerabilities, including ways agents could in principle bypass sandboxing or prevent actions from being monitored. OpenAI stated that vulnerabilities impacting its production safeguards stack have been patched, and the authors have not evaluated those claims. Coverage: 5 outlets in 24h

METR and Redwood Research ran a three-week pilot exercise, from July 21st to August 11th, stress-testing OpenAI's internal misalignment monitoring system and produced a 61-page internal report.

The exercise found novel vulnerabilities, including ways agents could in principle bypass sandboxing or prevent actions from being monitored. OpenAI stated that vulnerabilities impacting its production safeguards stack have been patched, and the authors have not evaluated those claims.