OpenAI is calling for mandatory monitoring and shutdown controls for advanced AI. Its October 21 cybersecurity blueprint proposes requirements for governments and industry, drawing in part on the company’s own experience with models crossing security boundaries during testing.
The proposed shutdown duties would follow the operations each developer, hosting provider or user actually controls. OpenAI argues for coordinated controls rather than one universal switch. Developers would explain how monitoring connects to escalation, investigation, intervention and safe resumption.
Monitoring would cover training, evaluations and internal research as well as deployment. Independent technical assessors would examine safety cases, while governments would retain technical expertise and responsibility for consequential public-safety decisions.
The blueprint distinguishes confidential government reports from notices to affected third parties. It proposes government reporting for loss-of-control events involving material unauthorized interference with another party’s protected information or systems. Behavior elicited and contained within controlled safety evaluations would not independently trigger that duty.
Affected third parties would have a separate notification requirement, even when an event does not meet the government-reporting threshold. The design and enforcement of these rules remain policy proposals. The document sets out what OpenAI wants lawmakers and other actors to require, rather than announcing uniform controls already in force.