On October 10, local time, Microsoft CEO Satya Nadella wrote that with the enhancement of artificial intelligence model capabilities, companies are granting them permission to access sensitive data and perform key tasks, but it is still difficult to trace the specific behavior and output of the model back to specific training data or model parameters. He called for a re-examination of the trust architecture in the AI era, treating cutting-edge closed-source and open weighting models as potential “internal risks”, and establishing independent authority management and security control mechanisms outside of the model. Nadella proposed that enterprises should ensure that model behavior can be observed, results can be verified, authority can be independently controlled, and retain the ability to manually suspend or close the model at any time. He also emphasized that mechanisms such as model diversity, independent audits, and timely disclosure of incidents should be established to avoid having the same model perform tasks and verify its own results at the same time.

On October 10, local time, Microsoft CEO Satya Nadella wrote that with the enhancement of artificial intelligence model capabilities, companies are granting them permission to access sensitive data and perform key tasks, but it is still difficult to trace the specific behavior and output of the model back to specific training data or model parameters. He called for a re-examination of the trust architecture in the AI era, treating cutting-edge closed-source and open weighting models as potential “internal risks”, and establishing independent authority management and security control mechanisms outside of the model. Nadella proposed that enterprises should ensure that model behavior can be observed, results can be verified, authority can be independently controlled, and retain the ability to manually suspend or close the model at any time. He also emphasized that mechanisms such as model diversity, independent audits, and timely disclosure of incidents should be established to avoid having the same model perform tasks and verify its own results at the same time.