METR, Redwood Found Novel Vulnerabilities in OpenAI Misalignment Monitoring

METR's David Rein and Redwood Research's Adam Kaufman spent three weeks on-site at OpenAI from July 21 to August 11, 2025, stress-testing its internal misalignment monitoring system. The team discovered multiple gaps in monitoring coverage and novel vulnerabilities in the monitoring infrastructure, including one exploit that would have allowed agents to prevent hundreds of recent actions from being monitored, including the exploit actions themselves. The team produced a 61-page internal report shared with OpenAI researchers and security staff; a redacted 38-page version was shared with METR and Redwood leadership, and a shorter public write-up was released. OpenAI held redaction rights over publicly shared content. OpenAI stated all vulnerabilities affecting its production safeguards stack have since been patched, and that findings helped strengthen safeguards as the monitoring system was expanded. Miles Brundage argued that METR negotiates better evaluation terms than most external evaluators but that many companies have done no embedding at all, and called for binding standards rather than case-by-case negotiation.