Microsoft CEO Satya Nadella: Treat AI as an internal risk, prepared to pull the plug at any time.

Beating AI News: Microsoft CEO Satya Nadella published a lengthy article advocating that AI models be treated as part of internal enterprise security risk management. Even non-malicious models can make mistakes or be exploited by attackers. Especially when models can access sensitive data and perform critical operations on behalf of users, enterprises cannot rely solely on security promises from model vendors. He called for separating models from the Harness (Agent execution framework) that schedules tasks. Which data a model can read and which tools it can call must be controlled by an external system. Each critical operation should leave tamper-proof, human-auditable records. He also demanded transparency in model chain-of-thought (the displayed reasoning process), while acknowledging that this content may not be reliable. The most stringent requirement is to assume models may be compromised from deployment and implement isolation measures. Authorized personnel must be able to pause or shut down an Agent at any time during task execution, equivalent to an "emergency brake". Models cannot audit their own behavior; enterprises must also continuously test for faults and attacks, and disclose incidents in a timely manner. Microsoft announced this month that its MXC isolation tool is now officially available on Windows. It restricts the files and networks Agents can access via external policies.