OpenAI built a 35-second kill switch for frontier AI

As AI shifts from answering prompts to taking autonomous actions, the nightmare scenario isn't just an incorrect answer—it's losing the ability to pull the plug.

OpenAI has introduced Control Room, a centralized system designed to monitor workloads across its research infrastructure and hit the brakes when safety risks emerge.

A 35-second emergency brake

Control Room connects directly to OpenAI's research inference workloads, tracking agentic sampling used in model training and evaluation, as well as data processing pipelines.

Working alongside OpenAI's misalignment monitoring tools, the system can halt model output or pause a running job within 35 seconds of an intervention decision.

When models act as autonomous agents running multi-step loops, waiting for a run to finish before auditing logs is no longer an option. A model drifting off-course can quickly generate corrupted synthetic data or trigger security concerns. Control Room ensures humans maintain a direct leash on these workloads.

Why it matters

We think this is a much-needed reality check for the frontier AI race. As autonomy increases, safety can't just be an after-the-fact red-teaming exercise; it has to be integrated into runtime infrastructure.

Establishing a 35-second hard cutoff sets a tangible engineering standard for cluster control. While Control Room is currently an internal mechanism that outside developers won't directly touch, it sends a clear signal to the industry: building more powerful autonomous models requires building an industrial-grade kill switch first.