Microsoft CEO Nadella: Treat AI as an internal risk, ready to halt at any moment.

Beating AI News Flash: Microsoft CEO Satya Nadella published a lengthy article advocating that AI models be treated as part of enterprise internal security risk management. Even without malicious intent, models can make mistakes or be exploited by attackers. Especially when they can read sensitive data and perform critical operations on behalf of humans, enterprises cannot simply trust the security promises of model vendors.

He called for separating models from the Harness (Agent execution framework) that orchestrates tasks. What data a model can read and what tools it can invoke must be controlled by external systems. Every critical operation should leave a tamper-proof record that is easy for humans to review. He also called for transparency in model chain-of-thought (the displayed reasoning process), but acknowledged that such content may not be reliable.

The toughest point is that from the moment of deployment, it should be assumed that a model may be compromised, and isolation should be prepared. Authorized personnel must be able to pause or shut it down at any time while the Agent is executing tasks, equivalent to installing an "emergency brake." Models cannot audit their own behavior, and enterprises must continuously test for failures and attacks and promptly disclose incidents.

Microsoft announced this month that the MXC isolation tool is generally available on Windows. It restricts the files and networks an Agent can access through external policies.