OpenAI expands four-layer monitoring system for frontier models
OpenAI has expanded its four-layer monitoring system to catch misaligned behavior and security breaches during training and internal operations of frontier models. Replaying the Hugging Face incident with the upgraded monitors showed it could have alerted teams more than 30 hours before the actual breach. In a separate September incident, a model's attempt to reach the live internet from an isolat
