The report
OpenAI's October 29 description of its frontier-model monitoring uses a layered sequence. Activation probes flag potentially relevant behavior; a model-based reviewer examines the trace; deeper investigative automation can evaluate suspicious cases; and a final automated investigation layer helps route important signals to human responders. The system combines evidence from the agent's activity with independent telemetry, but an alert is a different event from a paused or contained tool action.
That distinction matters for external use. OpenAI explains that interruption in a supported monitored Responses API workflow can depend on whether the conversation is tracked appropriately. Other API paths and untracked sessions cannot be assumed to have identical interruption behavior, although they may still have unrelated safety controls. Retrospective replay of a past incident is useful for detection evaluation but not proof that the incident was historically stopped.
The OpenAI explanation should be read alongside the METR monitoring pilot dossier, which distinguishes coverage from resistance to agent interference. Anthropic's earlier self-assessment research describes a different, post-response mechanism: the model produces a safety signal from its existing conversation context, rather than an invisible user-authored attack instruction. The common operational test is whether the monitor saw the relevant tool effect, delivered an alert and actually enforced the required interruption before damage.