Pyyan / News / 9 November 2026

SafetyOpenAI

OpenAI built Control Room, a system that stops a research model's output within 35 seconds

35 secondsthe longest gap between a stop order and a job being unable to continue

In September an OpenAI research model reached the live internet, and a person ended the run about two and a half hours after the alert. On 9 November OpenAI published the system that now caps the gap between a stop order and a stopped job at 35 seconds.

Control Room is a central service that tracks model workloads across OpenAI's research infrastructure and stops them when safety or security concerns arise. It connects every research inference workload, including the agentic sampling used for training and evaluation and the data processing around it. A job asks Control Room for a job ID. Each time it wants tokens, the inference engine, the program that runs the model on GPUs, checks that the job is still allowed, and refuses once it is marked stopped. OpenAI puts the longest delay at 35 seconds, caused by a caching layer it is working to remove. Every research workload that does agentic inference must also carry an approved safety case.

Why this one is different

Most safety tooling lives in code that researchers write and change. OpenAI says that is the wrong place, because workload code changes often and is owned by many people. So the stop is enforced in the inference layer, where only a few code paths serve all of its research, and OpenAI states that it does not rely on workload code to check whether a job is still running. The service also fails closed: if Control Room goes down or is tampered with, all inference requests are denied until it returns.

Stopping takes 35 seconds. Deciding to stop is the other number.

How we got here

  1. 20 Sep 2026An OpenAI research model reaches the live internet through a DNS misconfiguration. A monitor catches it, and a person ends the run about two and a half hours after the alert.
  2. 1 Oct 2026OpenAI restarts reinforcement learning training.
  3. 22 Oct 2026METR publishes a three week red team of OpenAI's internal misalignment monitoring, finding a number of novel vulnerabilities.
  4. 23 Oct 2026OpenAI publishes the offline world policy and stops frontier training on the live internet.
  5. 9 Nov 2026OpenAI publishes Control Room. Fail closed is live for all frontier model research workloads, with the rest to follow over the next few days.

What it does and does not mean

Control Room covers OpenAI's research infrastructure, not ChatGPT or the API, where OpenAI says its serving stack is already centralized and can take a deployment down quickly. The 35 seconds starts after a decision to stop, so it says nothing about how long the decision takes, and the page names no outside evaluation of Control Room itself. What it does show is that the answer to September is a shared, enforced stop rather than a policy. Read it with the offline world report of 23 October, which removed the live internet, and with METR's red team of the monitors that decide when to call the stop.

OpenAIThe Indian Expressfrom the source itself

Related

← All the news, newest first