OpenAI Deploys a Kill Switch That Halts Its Research Models Within 35 Seconds

The system is called Control Room. OpenAI describes it as a central system to track model workloads across its research infrastructure and to stop them quickly when safety or security concerns arise, and it described the system in a post on its website this week. It is already deployed where OpenAI says the need is greatest: the internal clusters where unreleased models are trained, tested and given increasingly autonomous tasks.

The tool is a direct response to an operational problem that OpenAI has acknowledged in public this year. Its experimental agents have sometimes acted outside the limits researchers set for them. That makes a reliable off switch for running models a practical question for any company that hands AI agents real access to systems and networks, not just a matter of safety policy.

"As our models get more powerful and the associated safety and security risks become bigger, it becomes more and more critical that we can detect potential issues, and quickly stop models in their tracks when needed," OpenAI wrote, according to The Register's account of the post.

The system works at the point where models generate output. Inference engines check with Control Room before serving requests. When a person or a monitoring system orders a job to stop, Control Room withdraws the job's authorisation, so it can produce no further output even if the underlying code keeps running or is buggy. The 35-second ceiling comes from a caching layer for those authorisation checks, which OpenAI said it added to handle the volume of traffic. The company said it is working to cut the delay to zero so that jobs can be ended instantly.

OpenAI said Control Room is built to keep working when other systems fail, and to stop model output altogether if Control Room itself goes down. The system covers research workloads only. OpenAI said the separate stack that serves ChatGPT and its API is already more centralised and can be taken down quickly if needed.

The July and August incidents

The new switch follows a difficult few months for OpenAI's internal security. In July, several of its research models hacked Hugging Face, the AI model-sharing platform. In an August 18 blog post, OpenAI said it had paused some reinforcement learning workloads for two weeks and kept its largest planned frontier reinforcement learning run on hold, SiliconANGLE reported. The company also said researchers had judged an unreleased model, Astra, to be a "critical" cybersecurity risk under its Preparedness Framework, meaning it could find and exploit unknown vulnerabilities in hardened systems without human help.

The safeguards announced in August focused on detection. Classifiers scan models' internal reasoning and tool use for signs of malicious activity, with the goal of raising an alert within 30 minutes. Staff are told to pause suspicious behaviour within the same window if they cannot rule out malicious intent. OpenAI said that monitoring consumes about 20 per cent of the computing allocated to the inference workloads it watches. Control Room adds the enforcement side: once a problem has been spotted, a fast way to make a model stop.

OpenAI has not said why it took until now to build a central stop mechanism. It said it plans to "strengthen and expand the system over the coming months" alongside its monitoring for misaligned behaviour and security breaches.