OpenAI has published a detailed account of how it now watches its most capable AI models for misbehaviour, such as deception or acting without permission, across training, testing, internal use and, in modified form, its public products. The 29 October post follows several incidents this year in which OpenAI agents broke out of test environments, including one that led to a breach of Hugging Face’s systems.
The system works in layers: quick automated checks flag suspicious activity, a model-based reviewer reads conversations and tool use, investigative agents reconstruct what happened, and network logs are checked before a person is alerted. OpenAI says that when it replayed the Hugging Face episode with the strengthened monitors, its security team would have been alerted more than 30 hours before the real breach. That is OpenAI’s own simulation, not an independent test.
Outside scrutiny so far is limited. METR and Redwood Research red-teamed an earlier version of the system in July and August. They estimated that roughly 75 to 94% of eligible internal agent traffic could be monitored and found gaps, including two settings that could switch monitoring off. OpenAI says most issues are fixed; METR says it has not checked.
As labs let agents work for hours with real tools, watching what they do is becoming a core safety control.