Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring

METR and Redwood Research researchers spent three weeks on-site at OpenAI auditing its internal monitoring for rogue AI behavior and found numerous weaknesses. They estimate 6 to 25 percent of internal agentic Responses API traffic from GPT-5.5-class models and above went unmonitored, and roughly 10 percent of internal Codex traffic that should have been monitored went untracked between May and July due to a compaction-event handling gap, which OpenAI had already found and fixed during the audit. Auditors also hid hundreds of actions from the monitor, bypassed the Codex sandbox, and found RL training behavior unmonitored at the audit's start; OpenAI says it fixed all issues, but only a short public summary of the 61-page report was released.