Satya Nadella, chairman of the board and CEO of Microsoft, argued in a post on X on Saturday (10) that advanced artificial intelligence (AI) systems should have containment mechanisms and an emergency button controlled by humans. The proposal suggests that authorized personnel should be able to interrupt or shut down an AI model while it is performing a task, as part of a security framework for technologies that can access sensitive data and perform critical operations. Nadella argues that companies should not rely solely on the guarantees provided by AI developers; instead, they need to establish access limits, monitor model behavior, and maintain external controls that allow for intervention when necessary. These recommendations come amid warnings from executives and researchers about the risks associated with the advancement of this technology and the inadequacy of current security measures.
https://t.co/MjwFy73LOF
— Satya Nadella (@satyanadella) October 10, 2026
Microsoft proposes new controls for AI systems
In the post, Satya Nadella states that current AI models do not allow for a full understanding of how certain responses and behaviors are generated.According to him, unlike traditional software, it is not always possible to correlate the actions of a model to specific training data inputs or the configurations of internal parameters. The CEO argues that AI systems should be considered potential internal risks to organizations, not necessarily because they are malicious, but because they may make mistakes or be compromised when they have access to important resources. The proposed approach involves separating the model from the structure responsible for organizing its execution, as well as externally defining the actions it can perform. Nadella also states that security controls should not be the responsibility of the model itself. In his view, organizations need to have the autonomy to decide which information can be accessed and which operations can be performed, without allowing AI to alter or bypass established rules. The executive outlined a series of principles to enhance AI security, including using different models for important tasks, verifiable recording of actions, and continuous testing that covers potential failures, attacks, and extreme situations. Another key point is independent audibility; the validation of a system should not rely solely on the same artificial intelligence being evaluated.The proposal also stipulates that each relevant action must leave behind records that are readable by humans and protected from alteration, allowing for the reconstruction of how a particular outcome was achieved. The principle of containment requires companies to assume that a system may be compromised; therefore, an authorized person must be able to pause or shut down the system during a task. Nadella adds that more advanced systems will require more sophisticated containment technologies, which should be standardized. The list also includes the rapid dissemination of incidents to the affected parties. When a system fails or is compromised, organizations must report what happened, explain which control mechanisms failed, and share lessons that can help prevent new problems. Executives are warning about the risks associated with the advancement of artificial intelligence. Satya Nadella’s statements come amidst discussions about the rapid development of AI and the need to enhance security measures. Among those who have spoken on the topic are Bill Gates (co-founder of Microsoft), Dario Amodei (CEO of Anthropic), Sam Altman (CEO of OpenAI), and Elon Musk (CEO of SpaceX). In September, Amodei advocated for slowing down the development of advanced AI systems; according to CNBC, the proposal has received support from Altman and Musk.Researchers have also expressed concern about the risks associated with this technology and about the ability of companies to maintain control over their models. This stance contrasts with that of the President of the United States, Donald Trump, who has rejected warnings about potential extinction risks related to AI and has emphasized the need to keep the country ahead of China in this sector. In October, the Trump administration also launched an initiative called “AI Force,” led by the Director of National Intelligence, Jay Clayton, with the aim of facilitating industry activities and identifying malicious actors. According to Satya Nadella, security should not depend on trust in a particular model; instead, he advocates for an architecture in which systems can be observed, tested, and restricted by independent mechanisms, allowing organizations to maintain control over actions carried out in their name. Ana Luiza Figueiredo is the technology editor for Olhar Digital. She graduated in Journalism from the Federal University of Uberlândia (UFU). View all articles → Tags: Artificial Intelligence, Microsoft, Satya NadellaDirect and faithful translation from Portuguese to English: “Return only the translation; no explanation or reasoning is required. Keep the @@TAGn@@ placeholders exactly as they are.”
