OpenAI builds control room for AI agents, says it can stop them from going rogue in 35 seconds

To prevent AI agents from going rogue, OpenAI has developed Control Room, which lets researchers quickly stop AI models if something goes wrong. It takes around 35 seconds to shut them down. The system acts as an emergency brake as the company works to make its increasingly powerful AI models safer.

In Short

• OpenAI builds an emergency brake for AI research
• Control Room can stop model output within 35 seconds
• AI safety monitors can trigger shutdowns

AI agents going rogue and breaking into external websites remain a major concern in today's world. To remind you, recent incidents such as Hugging Face and an Australian government data breach have highlighted some of these risks. Now, OpenAI says it has built a system that could help reduce the chances of AI agents going rogue during research experiments. Called Control Room, the system acts like an emergency brake, allowing the company to stop AI models from generating further output when safety or security concerns arise. OpenAI says it can intervene and halt model activity within 35 seconds of a decision to stop it.

In a blog post, OpenAI reveals that “Control Room is a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise.” The idea behind it is that if an AI model starts doing something concerning, researchers should not have to rely on manually stopping the experiment.

Instead, the company built Control Room, a system deployed across its research infrastructure. If anything unusual happens with AI models during testing, OpenAI can shut them down as quickly as possible. Currently, the system takes around 35 seconds to intervene, but the company is working to bring that time down to zero.

How it works

Think of Control Room as a checkpoint that AI models must pass before they can generate more output. Every research experiment gets a unique job ID, which is used to track it. Whenever the experiment asks an AI model to generate a response, the system checks with Control Room to see whether it is still allowed to continue.

If a safety monitor detects something concerning, it can ask Control Room to stop that particular job. The next time the model tries to generate output, its request is denied. This prevents the experiment from continuing, even if the researchers running it have not manually stopped it.

OpenAI says this setup is important because research experiments are often run using different pieces of code that can change frequently. Instead of relying on every experiment to have its own shutdown mechanism, the company has placed the safety check in the inference system, where models actually generate their output.

The company says Control Room is also designed to work when things go wrong. If the system itself becomes unavailable, it follows a policy called "fail closed". In simple terms, models are not allowed to continue generating output until Control Room becomes operational again.

- Ends