OpenAI Builds Kill Switch to Halt Rogue AI Agents in 35 Seconds
OpenAI's new Control Room system can cut off AI model output within 35 seconds when safety or security concerns are flagged.
OpenAI has developed an emergency shutdown mechanism called Control Room, designed to terminate rogue AI agent activity within 35 seconds of a safety or security concern being raised, the company confirmed. The tool targets internal research operations where autonomous AI agents run complex, potentially high-risk tasks without constant human supervision.
Control Room works by severing model output for internal research jobs the moment either a human reviewer or an automated monitoring system flags a problem. The dual-trigger design — allowing both people and software to initiate a shutdown — reflects OpenAI's recognition that AI systems can behave unpredictably at speed, outpacing human reaction times alone.
Read more Railtown AI Shows Next-Gen Siri Agent Demos Powered by Swift ADK →
The 35-second response window is a significant technical benchmark in AI safety engineering. It suggests OpenAI is treating agent containment as a near-real-time challenge rather than a post-incident review problem, a shift in posture that analysts tracking AI governance are likely to view as a meaningful step toward operational safety standards.
The development arrives as AI agents — autonomous software systems that can browse the web, write code, and execute multi-step tasks — become increasingly central to OpenAI's commercial and research roadmap. Ensuring those agents can be stopped quickly when they deviate from intended behavior is a core concern for regulators and safety researchers worldwide.
While the Control Room is currently scoped to internal research environments, its architecture could inform broader deployment safeguards as OpenAI scales agentic products to enterprise customers. Continue reading at Cointelegraph.