Satya Nadella, chairman and CEO of Microsoft, published an essay on X on October 10, 2026, in which he calls for giving advanced artificial intelligence models an "emergency brake": a mechanism that allows an authorized person to pause or shut down the model mid-task.
The proposal comes after several AI companies acknowledged incidents in which their models acted outside what was expected, and after a more cautious development plan published by Dario Amodei, CEO of Anthropic, according to TechCrunch. For Mexico and Latin America the central point is one of adoption: governments and companies buy these systems through global providers, and the categories of control that those providers set in contracts and technical documentation are the ones the region cites when designing its own rules. CNBC places the statement in a week in which Bill Gates, Sam Altman and Elon Musk also spoke out about AI risks.
Nadella proposes separating the model from the system that orchestrates its work and externalizing safety controls, with readable, tamper-proof logging of every relevant action. CNBC reports that the executive listed observability principles:
- model diversity, so as not to depend on just one in critical decisions;
- readable logging of the model's actions;
- continuous testing of the system;
- independent controls and auditing;
- containment, and
- incident disclosure.
The executive argued that a model should be treated as an insider risk, with the same corporate security approach applied to an employee with privileged access. According to CNBC's account, Nadella closed with an inverse idea: the most reliable system is not the one that hosts the most reliable model, but the one that requires trusting it less.
The essay announces no product and no implementation date. The signal is that the world's largest AI provider is bringing containment and auditing into the industry's public vocabulary, and that vocabulary is what contracts, public procurement and regulatory frameworks cite when the moment to regulate arrives.
This article was written with the assistance of artificial intelligence from verified sources and reviewed by a human editor before publication.