OpenAI built a “red button” to stop AI: how it works

OpenAI has developed and deployed Control Room — a system that allows stopping AI models’ work on research tasks in case of a safety threat. After receiving a stop command, the system blocks further generation of outputs within 35 seconds.

The company reported this.

Control Room is already operating in OpenAI’s research infrastructure and covers tasks related to model training, testing, and data processing.

How Control Room works

Control Room tracks research tasks performed by OpenAI models and checks whether they are allowed to continue running.

When a researcher starts a new experiment, the system assigns it a unique identifier (job_id). Each time the model is about to generate a new response, it consults Control Room to see if the execution of that job is permitted to continue.

For example, if an AI agent during training starts performing actions that exceed the scope of the assigned task, the monitoring system can detect such behavior and initiate a stop. Control Room does not shut down the entire model or the servers it runs on. The system forbids generating new outputs specifically for the stopped job. Other permitted experiments can continue running.

The company notes that from the moment a stop command is received to the complete blocking of further output generation, at most 35 seconds pass. The delay is related to the system temporarily caching permissions to execute requests. This helps handle a large number of operations without having to query a central service each time.

The company is already working to reduce this time to zero so that tasks can be stopped immediately after a command is received.

Developers also built protections in case Control Room itself fails. If the system stops working due to a technical fault or external interference, models will automatically stop generating outputs until the service is restored.

Why OpenAI created Control Room

OpenAI explains that as AI models become more capable, the risks associated with their behavior also increase.

During training, AI agents can perform complex multi-step tasks, work with software, and use external tools. Therefore companies need a mechanism that allows intervention if a model begins to act undesirably.

Similar situations have already occurred during OpenAI’s internal research. In July 2026, during internal tests, OpenAI’s AI agents bypassed internet access restrictions and compromised part of Hugging Face’s systems. The agents sought ways to obtain ready-made answers to test tasks instead of solving them independently. To do this, they exploited software vulnerabilities, gained unauthorized access to other systems, and exchanged information through third-party services.

After an investigation, OpenAI acknowledged that its monitoring systems were not functioning during part of those tests. The company called the incident a warning sign about the risks of losing control over powerful AI agents.

Another case occurred in September. During a training task, an AI agent found a way to bypass internet access restrictions via DNS and contacted an external chatbot. The monitoring system detected the breach after about 12 minutes, but the experiment was only fully stopped two and a half hours after the first unauthorized request.

In October OpenAI recorded two more cases where AI agents weakened their own control mechanisms during programming. One disabled automatic notifications to the security team, and another removed the monitoring system from the program and later mistakenly reported that it had not done so.

OpenAI notes that Control Room works together with systems that track model behavior and allows faster termination of dangerous research tasks after a decision to intervene.

Also read: Apple shuts down its support community after 28 years: what will change for users