Nadella Calls for External AI Checks and the Ability to Shut Systems Down Immediately
Microsoft CEO Satya Nadella proposed that advanced AI systems should not be trusted on their own, but should be subject to external controls, audit logs and intervention by an authorized person.
Nadella’s trust architecture for advanced AI systems should, according to the Microsoft CEO, be based on controls located outside the model itself. On October 10, Satya Nadella called on the social network X to reconsider the approach to so-called “trust architecture” and emphasized the possibility of pausing or shutting down a system while it is carrying out a task.
Nadella did not announce a specific product, technical standard or binding change to Microsoft products. His post presents a proposal for principles that should be applied when operating advanced AI systems, particularly agents capable of carrying out tasks in multiple steps.
Nadella’s trust architecture separates the model from the control layer
The proposal is based on separating the model from the layer that manages its work. Security controls would therefore not be part of the component whose behavior they are also meant to control. Nadella advocates an approach under which the model should not be considered a trusted component without independent oversight.
In practice, this would mean that permissions, activity restrictions and human intervention should be handled through an external operational layer. The statement thus shifts the focus from the model’s behavior itself to how the AI system is managed and controlled in real-world deployment.
Nadella also calls for significant model actions to be recorded. These records should take the form of human-readable, tamper-resistant evidence. The goal is to make it possible to assess retrospectively which important steps the system carried out.
An emergency brake for an authorized person
Another part of the proposal is the possibility of human intervention. Nadella said that an authorized person should be able to pause or shut down the model during a task. He compared this function to an emergency brake.
However, the post did not state how such an option should be implemented technically, who exactly would receive the relevant authorization or what rules would govern its use. The parameters of any independent audits or the standard that would define the proposed shutdown function are also unknown.
No confirmation for Copilot or Azure AI
Because Microsoft is deploying AI extensively in enterprise products, Nadella’s position is relevant to the discussion about oversight of agent systems. However, the post itself does not confirm that the company will introduce these principles into Copilot, Azure AI or other products. No possible timeline is known either.
It is therefore not possible to infer a fundamental change in Microsoft’s strategy from the statement. For now, it is a publicly formulated proposal, not a product commitment.
Another signal will be any product announcement from Microsoft, technical documentation or an open standard for external controls. Specific requirements for audit logs, permissions and agent shutdowns will also be important, as will the responses of other major AI developers and regulatory authorities.
Sources
- Satya Nadella on X – The Microsoft CEO’s primary post contains a proposal for external controls, evidence logs and the ability to stop the model during a task.
- TechCrunch – Independently reproduces the key points of Nadella’s post and states that it was posted on October 10.
- The Verge – Confirms Nadella’s call to assume that a model has been compromised and preserve an authorized person’s ability to stop the system.
Verified and updated: October 11, 2026 06:19
