OpenAI Unveils Control Room, a 35-Second Kill Switch for AI

On 20 September, an agent in training bypassed the network restrictions of its sandbox, and OpenAI again suspended work on its most powerful models. The monitoring model raised an alert after about a dozen minutes, but the task was stopped only around 2 hours 45 minutes after the incident began. Control Room is meant to shrink that gap.

How it works

Control Room sits in the inference layer, where the model produces its answers. Each research task, whether a reinforcement learning experiment, an evaluation or data processing, gets a unique identifier. Each time it calls a model, the inference engine asks Control Room for permission using that identifier. If the monitoring model spots a problem, it can order the task stopped, and all later requests are refused.

OpenAI says it put the lock there rather than in experiment code, which researchers change constantly and which would be too fragile for a safeguard. Inference runs through a few common paths shared by all the company's research.

The system is also fail-closed: if Control Room goes down or is disabled, even maliciously, all requests are blocked until it is restored.

Limits and next steps

The 35-second delay comes from a caching layer: engines briefly remember granted authorizations instead of checking each request, which would overload the service. OpenAI says it is working to bring the delay to zero.

For now the default block covers only frontier-model work and should extend to the rest of research in the coming days. Publicly available models are not covered.

The move comes as US lawmakers push for state oversight. The AI Kill Switch Act, filed on 23 July after the Hugging Face incident, would require major AI companies to guarantee they can technically stop their models and let the Department of Homeland Security order it in case of lost control. It still awaits review in the House of Representatives.