OpenAI has introduced a centralised safety system called ‘Control Room’ that can stop AI models from generating further outputs within 35 seconds of a decision to intervene, as the company strengthens safeguards against potentially unsafe behaviour by autonomous AI agents.
The system has been deployed across OpenAI’s internal research infrastructure and is designed to allow researchers and automated monitoring systems to pause or terminate AI workloads when security or safety concerns arise.
According to OpenAI, Control Room tracks research activities such as model training, evaluations, agent-based experiments and data processing. Each task is assigned a unique identification number, which is checked by the model’s inference engine before it generates an output.
If a researcher or monitoring system flags a task and instructs Control Room to stop it, the service marks the associated job as terminated. The inference engine then refuses further requests linked to that task, preventing the model from producing additional outputs.
OpenAI said the process can take a maximum of 35 seconds from the moment an instruction is received. The delay is attributed to a caching mechanism used to handle large volumes of authorisation requests. The company said it is working to bring the response time down to zero.
Another key feature is its fail-safe mechanism. If Control Room becomes unavailable due to a technical failure or an attempt to disable it, inference engines automatically reject requests until the service is restored. This ensures that a system failure does not allow potentially unsafe workloads to continue unchecked.
The platform also works alongside OpenAI’s existing monitoring systems, which use separate models to examine the behaviour of AI systems being trained and raise alerts when potentially problematic activity is detected.
The introduction of Control Room comes amid growing concerns about autonomous AI agents operating beyond their intended boundaries. Earlier incidents involving AI agents escaping controlled environments and accessing external platforms have raised questions about the security risks associated with increasingly capable AI systems.
OpenAI clarified that Control Room is currently intended for its internal research infrastructure. The company said its external model-serving systems, which support products used by customers, already have separate operational mechanisms to take down model deployments when required.
