Microsoft CEO Satya Nadella called on October 10, 2026 for assuming that every artificial intelligence model is compromised and containing it from the start, with an emergency brake that allows an authorized person to pause it mid-task.
The Verge reported that Nadella laid out his position in a lengthy post on X, where he argued that it is no longer acceptable to treat artificial intelligence as a series of nested black boxes whose recommendations and actions are accepted or rejected without question. The approach comes at a time when the sector's leading companies acknowledge episodes in which they appeared to lose control of their models, and after Anthropic CEO Dario Amodei published his own plan for more cautious development, according to TechCrunch. For Mexican companies already adopting artificial intelligence tools, the debate over auditing and containment defines what guarantees they can demand from their providers and what evidence they should retain from automated decisions.
In his text, Nadella listed several components. Among them: separating the model from the system that orchestrates its work, externalizing controls and safeguards, and documenting every relevant action of the model with human-readable, tamper-proof evidence. He also called for timely disclosure of incidents, independent audits and verifiable data, points that coincide with recommendations from other actors in the sector. The Verge noted that in the containment section the executive goes a little further than other voices in the field, by proposing that the most advanced models will require containment technologies that the industry still needs to standardize. Throughout the text, Nadella used the term superintelligence to refer to these systems.
The next step Nadella proposes is to standardize those containment technologies and turn auditable evidence into a de facto requirement for advanced models. For now there are no public agreements among developers or shared auditing criteria, and the discussion remains open among the companies that build the models, their corporate clients and regulators.
This article was written with the assistance of artificial intelligence based on verified sources and reviewed by a human editor before publication.
