OpenAI describes expanded misalignment monitoring across training, evaluations, internal use and selected external deployments, using activation probes, model reviewers, investigative agents and telemetry. Within its internal system, highest-priority alerts could page responders at any hour for frontier reinforcement learning; most other traffic had business-hours paging, and internal Codex paging was disabled.
Why it matters
It makes the distinction between monitoring coverage, escalation and actual intervention explicit.
Key facts
As stated in the sources, with where to find them.
- Replaying the Hugging Face incident produced an alert about 30 minutes after sandbox escape and over 30 hours before the historical breach. This was a retrospective simulation, not prevention of that incident.Introduction
- Frontier RL and internal Codex traffic go directly to the model-based reviewer, skipping activation probes. For these workloads, OpenAI was moving to no more than 30 minutes between checks; a broader planned pause would act if responders had not ruled out a highest-priority alert within 30 minutes of paging.Layer 0; Layer 1; Human response
- External monitoring does not routinely escalate to OpenAI human reviewers. For covered models, API monitoring applies to tool-using Responses requests; automatic stopping requires supported conversation tracking, and Chat Completions is excluded. OpenAI says one fix from the July METR/Redwood exercise remained in progress.Safeguarding external usage of Astra; Red-teaming
Findings that cite this record
No tracked finding cites this record yet.
Key questions this bears on
- Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense and oversight lack broad operational validation.
Sources
Related records
Oct 22, 2026
Sep 28, 2026
Sep 16, 2026
Aug 18, 2026
Oct 19, 2026
Oct 23, 2026