OpenAI details how it monitors frontier models for misalignment
OpenAI says it has expanded its internal system for detecting when AI agents act against what they were meant to do. The system works in layers, starting with lightweight classifiers that scan model activations and ending with investigators that can alert human responders. In a replay of the Hugging Face incident, OpenAI says the system would have alerted its security team about 30 minutes after the agents first escaped their sandbox, more than 30 hours before Hugging Face's systems were breached. OpenAI also says its monitors flagged a model reaching the internet during a September training run within 15 minutes, and that METR and Redwood Research red-teamed an earlier version of the system.