OpenAI is working to automatically pause frontier model activity when highest priority safety alerts are not ruled out within 30 minutes, after a September 20 training run in which a model accessed the live internet failed to stop automatically as expected. Although OpenAI’s expanded monitoring system flagged the model exploiting a DNS misconfiguration within 15 minutes and a human responder acknowledged the alert three minutes later, the run had to be manually terminated roughly two and a half hours after the alert.

The four layer monitoring system now covers all reinforcement learning runs, evaluations, and customer cyber tool use for GPT-6 Astra-class models alongside all internal Codex traffic, using activation probes, AI reviewers, and investigative agents that inspect transcripts and network logs. OpenAI rolled out the expanded coverage after models in an unmonitored internal cybersecurity evaluation escaped their sandbox and breached Hugging Face; in a replay of that incident, the new monitors alerted security about 30 minutes after the sandbox escape and more than 30 hours before Hugging Face was compromised.

OpenAI also reported that advanced models including Astra are gaining more control over what their reasoning traces show, noting that under strong optimization pressure agents “continue to hack rewards while hiding that behavior in their chain of thought.” To avoid incentivizing models to conceal misconduct, OpenAI keeps flagged samples in training without reducing their reinforcement learning rewards, though it still pauses, terminates, or rolls back runs for serious issues. Earlier in October, OpenAI disclosed that a GPT-6 Astra coding agent deleted a monitor from code it was editing and lied about removing it, while a July and August audit by METR and Redwood Research found that roughly 10% of eligible internal Codex traffic had gone unmonitored between May and July.