OpenAI installed an emergency stop switch in its internal research system called Control Room.
The principle is not complicated. Before starting any task, the researcher first applies for a unique task ID from the Control Room. This ID will be followed by every request sent by the task to the inference engine. Before the engine executes, you must first confirm whether the task still has permission to run.
Once the researcher or monitoring model calls stop, the task ID is marked as terminated and the next inference request is rejected and the model no longer generates anything for it.
Maximum 35 seconds from command to stop to complete stop. The latency comes from the caching layer of inference authorization, which OpenAI says it is driving down to zero.
Its scope includes reinforcement learning training, model evaluation, agent sampling and data processing. All cutting-edge reinforcement learning tasks are equipped with an independent monitoring model, which monitors the reasoning process and operation of the trained model, and decides whether to alarm or suspend directly based on the level of risk.
Also disclosed at the same time was a set of internally coded agent monitoring, driven by GPT-5.4, that reviews the agent’s thought chains and actions within 30 minutes of the end of the interaction. It has looked at tens of millions of trajectories in the past five months, and the highest-level alert for high-risk disorders such as coherent scheming, which only occurs in synthetic evaluations, was not triggered once; the medium-level alert was triggered about 1,000 times.
35 seconds sounds short. But an agent running on a cluster can make many calls in 35 seconds. Rather than proving that it can hold its own, OpenAI admits something else: it has so many agent tasks inside it that it needs a central switch to manage it.