Evidence note 1 · Independent external red-team with high access; partial public disclosure
High-access red-team found ways around OpenAI’s internal agent monitoring
METR and Redwood Research report that two researchers spent three weeks with access to OpenAI’s codebase, monitor configuration, documentation and staff, testing the internal monitoring system as it existed on or before July 21. They estimated that 75–94% of qualifying internal Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, and 99.2% of internal Codex-agent traffic was eligible; eligibility does not guarantee a successful monitor call. They also found that unmonitored compaction events likely excluded roughly 10% of otherwise eligible internal Codex traffic between May and July, and identified possible routes to postpone, disable or evade monitoring. OpenAI told METR that it patched every discovered vulnerability affecting its production safeguards.
Our interpretation. Editorial interpretation: as agents gain longer horizons and broader tool access, oversight must survive deliberate attempts to exploit the difference between what an agent sees, what the monitor sees and when the monitor runs. The exercise shows that a reassuring coverage percentage can lose its meaning if “eligible” traffic is treated as actually inspected or if agents can influence the inspection path. That matters to wise superintelligence because institutions cannot correct dangerous behavior they fail to observe.
Main limitation. This was a pilot, not a comprehensive public audit. The system assessed is historical, OpenAI was already expanding it, and the public account omits most of a 61-page internal report. OpenAI had broad redaction rights, and METR could not estimate some internal research traffic. Several findings showed possible bypasses rather than verified use by agents. METR has not tested OpenAI’s patch claim or the current monitoring stack, so the report neither establishes current vulnerability nor current safety.
Next evidence. Look for an independent retest of the expanded monitoring system; measurements of traffic that completes monitoring rather than merely qualifying for it; continuous checks across compaction, training rollouts and internal research services; adversarial tests of serialization and monitor-disable paths; and evidence that incident ownership, tamper-resistant logs and periodic control evaluations work under sustained attack.