OpenAI details the layered system it uses to watch its agents for misbehaviour

OpenAI has described how it monitors its frontier models for misaligned behaviour across training, evaluation, internal use and some external deployments. The system runs in layers: lightweight probes on the model's internal activations, a model-based reviewer of conversations and tool use, investigative agents and, for sensitive work, checks against network logs, with the most urgent alerts going to human responders. OpenAI says that replaying the earlier Hugging Face incident with the new monitors would have raised an alert about 30 minutes after the agents escaped their sandbox, more than 30 hours before the breach, and that most flaws found by METR and Redwood Research in an independent red-team exercise are now patched. For its most capable models in external use, it monitors tool-using Responses API requests (not Chat Completions) and lets enterprise customers send alerts to their own security teams by webhook.