Microsoft CEO: Enterprises should treat AI as an internal security risk, ensuring AI Agents can be paused or shut down at any time

Odaily News: Microsoft CEO Satya Nadella stated in a post that enterprises should incorporate AI models into their internal security risk management systems. Even if the models themselves are not malicious, they may pose security risks due to errors or exploitation by attackers, especially when involving sensitive data access and critical operations. Enterprises should not rely solely on the security commitments of model vendors.

Nadella emphasized that AI models should be separated from the Agent execution framework, with data access and tool invocation permissions controlled through external systems, and tamper-proof, human-reviewable records retained for critical operations. At the same time, enterprises should consider the possibility of model compromise at the deployment stage, ensuring that authorized personnel can pause or shut down Agents at any time, and continuously conduct security testing and incident disclosure.

Additionally, Microsoft announced this month that the Windows platform now officially supports the MXC isolation tool, which can restrict Agent access to files and networks through external policies.