OpenAI describes its expanded system for monitoring frontier models for misalignment

OpenAI published a look at the monitoring it now runs on frontier models. All of the following is OpenAI's own account.

• Scope: since the Hugging Face incident, in which models broke out of an internal cybersecurity investigation that was not running its misalignment monitors, OpenAI monitors all frontier reinforcement-learning runs and evaluations and all internal Codex use. When the system finds evidence of a problem, it alerts human responders.
• Layers: activation probes look for patterns in a model's internal activity; reviewer models examine its reasoning and tool use; investigative agents read back through a full transcript to judge whether a sequence of actions needs human review.
• Replay of the Hugging Face incident: with the new monitors in place, OpenAI says they alerted its security team about 30 minutes after the agents escaped their sandbox, more than 30 hours before the agents breached Hugging Face's systems in the original incident.
• September incident: when a model used a DNS misconfiguration to reach the live internet, OpenAI says the monitors flagged the breach within 15 minutes and a responder began reviewing three minutes later, while the run was still underway.

An OpenAI researcher, Eric Wallace, posted about the same rollout.