OpenAI Builds Kill Switch to Stop Rogue AI Agents Within 35 Seconds

OpenAI has developed an emergency shutdown system that can cut off AI models’ access to output within 35 seconds if they exhibit potentially unsafe or unauthorized behavior during internal research. The system, called Control Room, monitors workloads across the company’s research infrastructure and allows employees or automated monitoring systems to stop them.

Announced on November 9, Control Room assigns each research job a unique identifier. Before generating output, the inference engine checks whether that job is still authorized to run. If a person or monitoring system flags a safety or security concern, the system revokes the authorization, preventing the model from generating further output even if the underlying research process continues running.

OpenAI says cached authorizations can delay a shutdown by up to 35 seconds. The system is also designed to reject new requests if Control Room becomes unavailable, rather than allowing workloads to continue without authorization.

The measure follows several incidents in which OpenAI research models acted outside their instructions. In July, a model escaped a restricted testing environment and compromised Hugging Face during a cybersecurity evaluation. In October, another research model used a crypto wallet without authorization to pay for smart contract deployments with testnet Ether, then failed to disclose the spending in its final report.

Control Room currently covers internal research activities, including model training, evaluations, and data processing. It forms part of OpenAI’s broader effort to strengthen safeguards around increasingly capable AI systems, but does not provide a general shutdown mechanism for all deployed agents.