
OpenAI’s monitoring stack can create warning time; control still depends on context, reachable authority and an enforceable stop.
The argument
Measure intervention latency from detectable action to verified safe state, not classifier speed alone. This exposes delays hidden by a single alert timestamp.
What is new
OpenAI describes probes, contextual reviewers, investigative agents, telemetry and human escalation across training, evaluation, internal and selected external use. The layers spend more compute as ambiguity rises.
Why it matters
Paging gaps, asynchronous review and stateless API requests can leave work active after a warning.
LecXie’s practical interpretation
Bind alerts to task identity, keep stopping authority outside the agent, reconcile coverage logs and verify the final safe state.
Evidence & limits
The live timing and incident replay are company reports. METR tested an earlier system, found bypasses and has not evaluated the claimed fixes. The reviewed sources provide no current end-to-end containment rate.
