Microsoft's head suggested considering AI models as a potential internal risk for companies

Satya Nadella, head of Microsoft, wrote on X about the risks of using powerful artificial intelligence models. In his view, companies need to rethink their approach to AI safety and not rely solely on developers' promises.

Unlike traditional software, the operation of AI models is difficult to trace back to specific data or settings. At the same time, such systems gain access to confidential information and can perform important tasks on behalf of employees.

Nadella proposed treating advanced AI models as potential internal risks. This does not mean that they intentionally cause harm, but any system with broad authority can make a mistake or become a target for an attack.

To enhance security, he outlined several principles: limiting access to data, logging all important actions, regular testing, independent verification of results, and the ability to stop the model at any time. He also emphasized the importance of disclosing incident information so that companies can learn from others' mistakes.

According to Nadella, it is not enough to simply trust AI responses or have one model check another. Control must operate independently of the model itself, and the company must independently determine which actions are allowed.

The reliability of AI depends less on trust in a specific model and more on a control system that limits its capabilities and allows for the verification of actions.