OpenAI details a four-layer system to detect misaligned behavior in its frontier models

OpenAI described an expanded monitoring stack, from activation probes to investigative agents and network telemetry, covering training, evaluation, internal use and, in modified form, external deployment. It says a replay of the Hugging Face incident would have alerted security over 30 hours before the breach; METR and Redwood Research red-teamed the system.