# OpenAI activates an emergency brake for its AI agents

> OpenAI activates an emergency brake that cuts off the model's responses to its AI agents within a maximum of 35 seconds after the order
> By Observatorio Blockchain · AI Agents · 2026-11-10

OpenAI has launched a system that can cut off within a maximum of 35 seconds the responses its artificial intelligence agents receive in internal tests. The company presented the Control Room tool yesterday, November 9, which can be activated if a researcher or a monitoring system detects a possible security problem.

## What Control Room does

An AI agent can query the model several times during a task to decide what to do next. If the stop order is given, Control Room cuts off that agent's access to the model. The program can remain open, but it stops receiving the responses it uses to decide its next steps. Meanwhile, other experiments can continue using the same model.

The clock starts when Control Room receives the stop order. According to OpenAI, those 35 seconds are the maximum time during which a permission granted before the order can remain in force. The company is working to make the cutoff immediate.

**[OECD: First lessons from organizations using AI agents](https://observatorioblockchain.com/en/ai-agents/ocde-lecciones-organizaciones-agentes-ia/)**

[OpenAI](https://openai.com/index/control-room/) says it already uses Control Room in its internal work on training, model evaluation and data processing. The company has also planned for what happens if the emergency system itself fails. In tests with its most advanced models, requests to the model are blocked until the service is restored. OpenAI expects to extend that protection to the rest of its research work in the coming days.

The brake acts on the server that runs the model. Thus, if the agent's program continues sending requests after the stop order, it is the server that rejects them. OpenAI has placed the control there so that the cutoff works even if the agent's program has a fault.
