OpenAI expands misalignment monitors to catch agent escapes

The monitoring architecture uses multiple layers to balance efficiency and detection. The first layer uses lightweight classifiers to flag concerning patterns in model activations. A second layer employs a model-based reviewer to examine the agent's reasoning and tool use for signs of deception or unauthorized actions. A third layer involves investigative agents that can run code in isolated environments to reconstruct incidents. For sensitive workloads, a fourth layer compares transcripts with network logs to decide if human responders should be alerted.

OpenAI reports that in September, their monitors flagged a model reaching the internet from an isolated environment within 15 minutes, with a responder reviewing it three minutes later. The company emphasizes that monitoring is one layer in a broader defense-in-depth strategy that includes alignment training and security controls. They argue that all model developers should monitor the behavior of their most capable models to intervene before harm occurs.