OpenAI Control Room: Stopping research model output within 35 seconds

OpenAI introduced Control Room on November 9 to track AI jobs in its research infrastructure and stop model output when intervention is needed. The company reports a maximum delay of 35 seconds after a stop instruction reaches the system.

Dji Banner Inside Article - 728x90

Central authorization for model requests

Training experiments, evaluations and data-processing jobs receive a job identifier. Before generating output, the inference engine checks whether that job remains authorized. If a person or monitoring system stops it, further requests are refused. Enforcement therefore does not depend on the research code shutting itself down.

What the 35 seconds measures

The interval begins with the stop instruction, not the first appearance of a dangerous behavior. Authorization caching, used to handle traffic volume, explains the delay. OpenAI wants to reduce it to zero. A stopped research process can remain alive: what ends is its access to new model output. Other authorized jobs can continue using the same engine.

Output is denied if the service fails

The design also rejects inference requests when Control Room becomes unavailable. OpenAI says this behavior is deployed for frontier-model research, with remaining research workloads following over the next few days. The announcement concerns research infrastructure rather than a new consumer ChatGPT setting.

Part of a broader safety approach

OpenAI’s October 29 account describes behavior monitoring through model-based reviews, deeper investigations and human response when necessary. Some critical alerts can pause affected work before a responder acts. Control Room enforces the stop decision that follows detection.

The company’s September 28 guidance describes an ambition to build structured safety cases: evidence-based arguments about the risks of frontier training. Alignment, containment and monitoring are complementary safeguards. A stop mechanism alone therefore does not establish that every risk has been eliminated. The timing and deployment claims above are OpenAI’s statements, not independently measured test results.

Sources: OpenAI Control Room, OpenAI monitoring overview, OpenAI safety-case guidance, Unite.AI coverage.