Microsoft CEO Satya Nadella has proposed a new trust architecture for frontier AI systems, centered on an emergency brake feature. This mechanism would allow authorized personnel to pause or terminate an AI model while it is in the middle of executing a task.
The proposal addresses the risks associated with agentic systems, which can perform sequences of actions independently. Nadella suggests that advanced AI should be managed under the assumption that it is already compromised, similar to protocols used for insider threats.
The initiative has received internal support, with Microsoft President Brad Smith publicly endorsing the concept in September 2026. This move aligns with broader industry discussions regarding safety standards for complex AI systems and the potential for loss-of-control risks.
While the proposal aims to establish Microsoft as a responsible steward of technology, it leaves several operational questions unanswered. Specifically, the criteria for authorized personnel and the standards for independent audits remain undefined.
