What the report says
OpenAI describes expanded monitoring across frontier training, evaluations, internal use, and specific external deployments. Its system uses several screening and investigation stages, with human response and some automatic pauses. These are OpenAI’s reported controls and results, rather than tests performed by Tinker & Tokens. [1]
Detection and intervention are separate
For models covered by this system, the API guide says Responses requests are monitored. Automatic stopping depends on persisted reasoning, WebSockets, or OpenAI compaction. Without those mechanisms, configured webhooks can receive alerts while the conversation continues. Registering an alert webhook adds no stopping capability by itself. Chat Completions requests are outside this monitoring system; other safety checks still apply. [2]
Monitoring runs asynchronously, can miss issues or flag legitimate work, and does not undo earlier actions. A registered safety block is not proof that execution has already stopped. The guide tells developers to stop further actions after a blocked request and avoid automatic retries of that workflow. [2]
Dates and scope
The report page and current OpenAI research listing show October 29; an indexed search result shows October 28. We have not established its first public hour. [1][3]
Our take
For experienced technicians building agents, map each consequential action to an approval, an audit record, and a way to stop further work. Check what actually changed after an alert. Additional monitoring is useful only alongside those controls and an operator who can respond; do not assume an alert reverses a completed action. We have not tested this system.