OpenAI hosts embedded METR–Redwood audit of its internal agent monitoring
Where it stands: OpenAI said it invited METR and Redwood Research on-premises for three weeks in summer 2026, with two employees given deep access to evaluate its monitoring systems for misalignment by internal agents. It said findings helped strengthen safeguards, a brief METR write-up is public, and future third-party collaborations are planned. It is the first record of a frontier lab hosting an embedded third-party audit of its own internal-agent monitoring; watch the write-up's findings and whether other labs follow.
Latest
- 2026-10-22 opened — OpenAI announced that in summer 2026 it invited METR and Redwood Research, two employees on-premises for three weeks, to evaluate with deep infrastructure access the systems OpenAI uses to monitor for misalignment from internal agents; findings "helped strengthen safeguards," a brief METR write-up is public, and future third-party collaborations are planned (x.com/OpenAI). Debut thread: first record of a frontier lab hosting an embedded third-party audit of its own internal-agent monitoring. Watch: the write-up's content and any disclosed findings, replication by other labs, follow-on METR collaborations. Headline: OpenAI gives METR and Redwood an embedded audit of its internal agent monitoring.