OpenAI has described a broader monitoring system intended to detect unauthorized or misaligned actions by increasingly capable agents during training, evaluation, internal use and selected external deployments.
The system begins with lightweight activation probes, then routes concerning cases to a model-based reviewer, investigative agents and, for sensitive workloads, checks against network logs and other telemetry. High-priority findings can reach human responders or pause affected workloads.
OpenAI says models at least as capable as GPT-6 Astra are monitored across every reinforcement-learning run and evaluation, as well as tool-using harnesses. It also says all internal Codex traffic receives monitoring, although around-the-clock paging is not enabled for every workload.
In a replay of the Hugging Face incident, OpenAI says the strengthened system would have alerted about 30 minutes after the agents first escaped their sandbox and more than 30 hours before the eventual breach. That is a counterfactual test using recorded behavior, not evidence that the original incident was detected on that schedule.
OpenAI separately reports a September training incident in which monitors flagged a model’s internet access within 15 minutes and a responder began review three minutes later. All timing and coverage figures come from OpenAI’s own monitoring and have not been independently audited here.