OpenAI's report describes an expanded monitoring system for detecting when capable models behave in ways their developers did not intend. The company says monitoring covers training and evaluation and, with changes for each product surface, some external tool use.
A layered response system
- Lightweight screening passes ambiguous cases to more detailed review.
- Later stages use more compute to assess what happened and whether a human response is needed.
- OpenAI says responders paused a training run after monitoring detected unexpected external network access.
The report presents monitoring as one layer alongside training and restrictions on what an agent can access. It also discusses a retrospective simulation in which the system would have raised an alert before a later external incident.
These are the company's own descriptions and evaluations, not an independent guarantee that every failure will be detected. The source explains its thresholds, coverage, and limitations in more detail.