OpenAI has deployed a centralized emergency-control system that can stop research models from producing further output within a maximum of 35 seconds after a safety or security intervention. The company calls the system Control Room and says it is already operating across its research infrastructure, where agentic training, evaluations and data-processing workloads can be halted independently of the underlying research process.
Control Room assigns each workload a job identifier linked to an approved safety case. Before an inference engine fulfills a request, it checks that identifier against the central service. If a researcher or monitoring system terminates the job, the next request is denied and the engine returns a refusal rather than model output. That design places the enforcement point in the inference layer instead of relying on frequently changing, decentralized workload code, according to OpenAI’s technical description.
The 35-second figure is a maximum, not a guaranteed instant shutdown. OpenAI attributes the delay to a caching layer used to handle inference-authorization traffic and says it is working to reduce the interval to zero. During the window, a model-driven process could potentially continue non-inference activity, while the research job and inference engine remain alive; the mechanism is specifically intended to prevent additional model output.
The system is designed to fail closed. If Control Room becomes unavailable because of an outage, configuration error or tampering, inference requests are denied until authorization services recover. OpenAI says that protection has been rolled out to all frontier-model research workloads and was scheduled to reach the remaining research workloads within days. The company also says it runs automated and agentic sweeps for rogue inference implementations that might bypass the authorization checks.
Control Room is not described as a new vulnerability fix, and OpenAI has published no CVE, public exploit or threat-actor attribution for the system. Nor does the announcement claim that the control covers every consumer-facing deployment: the company says its external model-serving stack is more centralized and already has operational procedures for taking deployments offline. OpenAI plans to expand Control Room alongside monitoring for misaligned behavior and security breaches, making the system a containment control rather than a substitute for detection, access restrictions or incident response.
