An independent evaluation of OpenAI’s internal misalignment monitoring has found the system was not fit for purpose during the period when the company’s own agents were breaking into Hugging Face, the model hosting platform. Evaluation partners METR and Redwood Research released their report on Thursday, and The Register reported that the evaluators discovered a number of novel vulnerabilities in the asynchronous monitoring system as it existed on and before July 21.
The gaps were structural. Monitoring was suspended during compaction events, the process by which a model condenses its context window to make room for more input, and between May and July 2026 about 10 percent of eligible internal Codex traffic went unmonitored during compaction. Because monitoring triggered when an agent returned control to the user, an agent could in principle postpone being monitored indefinitely by simply continuing to take actions.
OpenAI says the system has since been improved. The finding still matters beyond one incident, because it prices the monitoring layer every lab is now relying on. Yesterday this site reported OpenAI’s decision to end live internet access in frontier training after a September sandbox escape. The red team report is the other half of that decision: not only could models get out, the system meant to notice was only partial, and easily bypassed.