The next time you pick up a carton of milk, consider this: It has likely passed through more safety inspections than the AI system writing your emails, screening your job application or negotiating with your bank. That, according to Microsoft President Brad Smith, is the problem.
In an interview with ITV last week, Smith said that AI is less regulated than milk and washing machines. He proposed that governments adopt established production safety practices and that AI companies build “kill switches,” a kind of emergency brake that would allow operators to shut down a system that begins behaving in an unsafe or unintended way.
Smith’s comments come amid growing public concern that voluntary industry commitments to safety — a form of toothless self-regulation — are not enough to manage systems that can influence decisions, handle sensitive data and, in the wrong hands, enable cyberattacks. Recent cases of so-called “rogue AI agents” escaping their environments, such as the summer’s infamous Hugging Face hack and attempted breaches of Australian and US government websites, have drawn scrutiny to AI’s alleged “existential risk” and sparked calls for a pause.
More from CNET
On Saturday, Microsoft CEO Satya Nadella echoed Smith’s sentiments in a lengthy X post, saying that the biggest companies lack a “mechanistic understanding” of the tools they deploy. Nadella also argued for slower development, incident reporting and independent review.
“We must build contained systems whose behavior we can observe, limits we can test and actions we can always contain,” Nadella wrote.
Microsoft’s statements come less than two weeks after top AI executives, including Nadella, met with President Donald Trump, who announced his proposal to rename AI as “super intelligence.“
Several Silicon Valley leaders — including Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang and xAI founder Elon Musk — signed the White House Accord on Super Intelligence, pledging to establish internal committees to oversee development and to work with independent auditors. (Though Microsoft doesn’t appear to have signed Trump’s voluntary accord, Nadella notably uses the term “super intelligence” in his recent X post.)
Smith’s and Nadella’s comments place Microsoft closer to Anthropic’s camp of “caution,” which calls for stronger AI oversight. Meanwhile, OpenAI’s Altman said that we should all just “accept some bad things happening” for the benefit of AI, whereas Zuckerberg and Huang continue to double down on their rejection of an industry-wide, coordinated slowdown in AI development.
Wherever they stand on the regulation debate, AI leaders all stand to benefit in some way from engaging in the discussion on safety guardrails. For example, Microsoft recently released Microsoft Execution Containers, an enterprise security feature that protects devices by restricting the files and programs AI models can access and providing a built-in kill switch to instantly shut them down if needed.
A Microsoft representative declined to make any further comment.
The ‘rogue AI’ narrative and Big Tech regulation
Despite increasingly public safety threats, it’s unlikely that the Trump administration will back regulation across the entire AI industry. The president has declared AI safety concerns a hoax, even as insiders report that more nuanced debates are unfolding behind closed doors. Last month, Trump posted on Truth Social that the only guardrail the AI industry needs is a “strong and smart president.”
In December 2025, Trump issued an executive order designed to prevent states from passing “cumbersome legislation” regulating AI model development, passing the buck to his own administration. The president continues to post online about maintaining a low-regulation environment for the biggest AI companies.
Many prominent AI critics argue that the entire regulatory framework is flawed from the start because it lets AI companies define the security threats and respond on their own terms. Distributed AI Research Institute founder Timnit Gebru has argued that the “rogue AI” narrative surrounding incidents like Hugging Face and access to government websites shifts attention away from basic corporate failures.
“This is not models going rogue,” Gebru told Democracy Now during an interview on Oct. 1. “This is gross incompetence and negligence by a particular company.”
Gebru said that people need to understand that these fallible machines don’t act autonomously with their own agency. “We need to hold corporations liable for unleashing these kinds of error-prone models into the world,” she said.
According to Gebru, the focus on so-called runaway AI and the push for voluntary self-regulation are ways for the industry to avoid accountability under existing law. Rather than face liability for unsafe security practices occurring now, AI companies redirect the conversation to an unprecedented danger likely to occur in the future.
The real danger is regulatory capture, where AI companies shape the rules that are supposed to constrain them. In this way, the public is asked to trust the morality of the very companies that build and profit from AI.