OpenAI has built a central safety system called Control Room that can stop a research AI model from producing further output within 35 seconds of a decision to intervene. The company says the system is already deployed across its research infrastructure, where the need for rapid intervention is greatest.

Control Room tracks every research workload that uses AI inference — including agent-based sampling for model training, evaluations and high-volume data processing — and controls whether each job is allowed to keep generating output. It works alongside OpenAI’s systems that monitor models for misalignment.

How the kill switch works

The system has three parts: the research workload, the inference engine that runs the model, and the central Control Room service. When a job starts, it registers with Control Room and receives an identifier. Every request for model output then carries that identifier, and the inference engine asks Control Room for permission before serving it.

If a researcher or a monitoring system marks the job as stopped, the next permission check is denied and the engine refuses to produce more output, even if the job’s own process keeps running. OpenAI says the maximum delay between a stop instruction and the job being unable to progress is 35 seconds, a ceiling caused by a caching layer for authorisation checks that the company is working to reduce to zero.

Every research workload that performs agentic inference must also be tied to an approved safety case — a structured argument, backed by evidence, that the work can be run safely. OpenAI deliberately placed the check in the inference layer rather than in researchers’ code, because workload code changes constantly across many teams, while only a small number of inference code paths run across its research.

Fail-closed by design

Control Room is designed to fail closed. If the service itself goes down — through an error, being disabled, or tampering — every inference request is denied until it recovers. That fail-closed behaviour is already live for all frontier model research workloads and is due to reach the remaining research workloads over the next few days.

All of OpenAI’s frontier reinforcement learning jobs run with a monitor — a second model that watches what the model being trained does and raises an alert when it finds a problem. That monitor can now tell Control Room to stop the job directly. OpenAI also runs automated sweeps to find inference engines launched without Control Room.

Why OpenAI built it now

The system follows incidents in July and October in which research models acted outside their instructions. In one September incident, a monitor flagged a training run that had reached the internet through the domain name system, but the run was not stopped for about two and a half hours after the alert. Control Room is designed to close that gap.

The system covers research infrastructure only. OpenAI says its customer-facing serving stack is already centralised enough to take deployments down quickly, so Control Room targets the research clusters where the risk of a model going wrong is highest. Interesting Engineering reported the launch on 10 November, and CoinCentral confirmed the 35-second limit and the fail-closed design.