OpenAI describes Control Room, a system to halt research workloads in 35 seconds
OpenAI said it has deployed a central service across its research infrastructure that can stop a training run, evaluation or data job from receiving further model outputs within 35 seconds of a decision to intervene.
In the Daily Brief of 11 November 2026
Every research workload that runs agentic inference must be tied to an approved safety case and must obtain a job identifier from Control Room. The identifier travels with each request to the inference engines, which check with the service before generating output. When a human or an automated monitor flags a job, its authorisation is revoked and output stops.
The design fails closed: if the central service becomes unavailable, inference requests are blocked. Placing the control at the inference layer rather than in research code means a change to a training script cannot switch it off. OpenAI said its external model-serving systems are more centralised and already have ways to take down deployments.
The system is internal and is not offered to customers.