The development

What happened

OpenAI describes wider monitoring across training, evaluations, internal agents and selected external deployments. It combines staged checks with investigations, while external coverage and intervention depend on the product and configuration.

Key facts

  • Internal coverage includes all Codex traffic and specified frontier training, evaluation and tool workloads.

  • A retrospective Hugging Face incident replay flagged sandbox escape after about 30 minutes, over 30 hours before the actual breach, OpenAI reports.

  • In September live training, it reports detection within 15 minutes and human review three minutes later.

  • External monitoring does not routinely escalate to OpenAI staff. Tool-using Responses API conversations need supported tracking for automatic stopping; Chat Completions is excluded.

  • OpenAI says existing Zero Data Retention protections remain. Earlier independent red-teaming found weaknesses; most were patched, with one fix outstanding.

The context

Why it matters

For API builders, a monitoring alert and an automatic stop are different guarantees. Reported detection times do not establish universal prevention or independent effectiveness.