Three-week on-site audit of OpenAI finds major gaps in AI safety monitoring

Two researchers from METR and Redwood Research spent three weeks on-site at OpenAI auditing its internal monitoring system for AI loss-of-control behavior and found multiple weak points. According to METR's estimate, during the audit period 6% to 25% of internal agentic Responses API traffic from GPT-5.5-level and above models was entirely uncovered by monitoring; about 10% of internal Codex traffic that should have been monitored was not tracked between May and July due to a defect in handling compaction events (context summarization), which OpenAI independently discovered and fixed during the audit, and METR said both estimates have large uncertainty. The auditors also found methods to hide hundreds of operations from the monitor and to bypass the Codex sandbox, and judged that specially crafted inputs would likely crash the monitor entirely; at the start of the audit, model behavior during RL training was not monitored.