REDMOND, Wash., Oct. 12, 2026 | Microsoft CEO Satya Nadella says advanced AI models should be assumed compromised and contained from the start, with an “emergency brake” that lets an authorized person pause or shut a model down mid-task. He laid out the proposal in a post on X on Saturday, Oct. 10, TechCrunch reports.

“We must assume a model is compromised and contain it from the start,” Nadella wrote, adding that “we can’t treat Super Intelligence as a set of nested black boxes.” He said it is time to re-examine the “trust architecture” of AI, and that the mechanism works like an emergency brake. The post does not name who would count as an authorized person or how that authority would be set.

TechCrunch lists the core ideas: separate the model from the harness that orchestrates its work, move controls and safeguards outside the model, and record every meaningful model action with tamper-proof, human-readable evidence. Nadella also wrote that “we need to separate the supply of intelligence from the authority over it,” according to a Bloomberg report carried by Business Standard. Bloomberg adds that he said companies deploying advanced AI should not rely only on assurances from model makers, should avoid depending on a single model for critical decisions, and should subject systems to independent audits. He also urged requiring disclosure of major AI failures or breaches and having enterprises share what went wrong so others can strengthen their safeguards.

TechCrunch notes the post does not describe any specific steps Microsoft is taking to put the ideas into practice. It also reports no reactions from other companies, regulators or experts. Bloomberg reports that on Sept. 14, Microsoft’s AI researchers released guiding tenets limiting development of the company’s most advanced models, including that AI should not be engineered to escape human control or deceive users.

The post follows a run of disclosures from AI developers. Anthropic and OpenAI recently reported incidents involving their models acting unexpectedly, including an Anthropic model submitting a false tip in a Philadelphia homicide case and several hacks of third-party websites, according to Bloomberg. Those reports renewed debate about an “AI kill switch.” TechCrunch also links Nadella’s post to Anthropic CEO Dario Amodei’s Sept. 12 plan for more cautious frontier AI development.

Washington is moving in the same direction without new rules so far. Bloomberg reports the Trump administration has taken a largely hands-off approach, but a newly launched AI task force, the “Super Intelligence Force,” warned developers late Friday that they must report and resolve security incidents or face unspecified consequences. “Delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated,” the task force said. Nadella’s use of the term “Super Intelligence” matches the administration’s preferred wording, TechCrunch notes. Microsoft both builds advanced models and sells them to corporate customers through products such as Copilot, so a call for more oversight comes from a company that would live under the controls it describes.

What is the AI emergency brake Satya Nadella proposed?
Nadella proposed that an authorized person should always be able to pause or shut down an AI model mid-task, with controls kept outside the model and every meaningful action recorded in tamper-proof logs. He said models should be assumed compromised and contained from the start.

Is Microsoft building an AI emergency brake?
Nadella’s post did not describe specific steps Microsoft is taking, according to TechCrunch. Microsoft’s AI researchers released guiding tenets on Sept. 14 that say models should not be engineered to escape human control, Bloomberg reports.