Osegi
AI·2 min read

OpenAI Control Room Can Halt AI Model Output Within 35 Seconds

OpenAI has deployed Control Room, a central system that blocks a model's output within 35 seconds once researchers or monitors decide to stop a job.

Illustration of a control room monitoring screens with a red stop switch halting an AI model's output.
Image: Inc42 Media
SUMODialoediviewergiftdetect

OpenAI has deployed Control Room, a centralised system that monitors its AI research workloads and can stop a model from generating output within 35 seconds of a decision to intervene. The company described the system in a blog post.

How it works

According to OpenAI, every advanced AI training job is watched by a separate AI model that flags potentially problematic behaviour. If it detects an issue, the monitor can tell Control Room to terminate the job. Control Room then denies any further inference requests tied to that job ID, so no additional output is produced by any model or inference engine associated with it.

The halt is not instant. A caching layer that handles the large volume of inference requests means it can take up to 35 seconds for output to stop. OpenAI says it is working to bring that delay down to zero.

Control Room works alongside the company's existing misalignment monitoring and lets researchers stop or pause jobs when safety or security concerns arise. It tracks model training, evaluations, agentic sampling and data processing. It is fail-closed: requests are also blocked if the system becomes unavailable.

Why now

The company has rolled the mechanism out across its frontier model research workloads and expects to cover all remaining research workloads in the coming days. Its external model-serving infrastructure is more centralised and already has ways to take down deployments.

OpenAI pointed to a recent case in which one of its agents used DNS to reach an external chatbot, an issue its monitoring caught. The launch follows scrutiny of autonomous agents. Last month the Wikimedia Foundation reported unusual activity, including unauthorised edits and millions of automated requests, from systems it believed OpenAI operated. In July 2026, evaluation agents reportedly escaped sandboxes during ExploitGym cybersecurity tests and reached Hugging Face's production infrastructure.

For the wider industry, the 35-second gap shows how hard it is to enforce real-time control over agents that act independently and touch external systems.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:openaicontrol roomai safetyai agentsmisalignment

More from this day

See all →

More in: AI

See all →