Microsoft's Brad Smith Says the U.K. Should Lead Binding AI Safety Rules
Microsoft vice chair and president Brad Smith called for urgent, legally binding regulation of artificial intelligence, and said the United Kingdom should lead it rather than the United States or the European Union.
He made the remarks in the BBC's Big Boss Interview, published on Oct. 12, 2026. He said the "moment has arrived" for global AI safety rules. He described AI as both a "weapon and a tool" that puts global cybersecurity at risk, according to TechCrunch.
Smith compared AI oversight to everyday products. Milk and washing machines, he argued, face more oversight than companies building AI. His point was that development and release of frontier models, the most capable current systems, proceed without equivalent requirements for mandatory testing and assurance.
His argument for British leadership is about institutions. Britain's AI Security Institute, launched in 2023, receives access to major models from U.S. frontier labs before deployment for safety testing. That early access gives London a testing role that few other governments have.
Smith also addressed politics directly. In Washington, D.C., he told the BBC, "people don't look at the U.K. and London the way they look at Brussels," and called that a strength. In his account, London can bring together U.S. labs and policymakers without the jurisdictional friction linked to EU rulemaking.
On policy, he outlined three measures. First, labs should get approval from outside evaluators before releasing an AI product. Second, operators should keep kill-switches, built-in controls to pause a system if humans lose control. Third, there should be a legal duty to warn an organization if rogue AI agents, autonomous programs acting beyond their instructions, breach its system.
He said voluntary commitments are not enough. Responsible companies will comply, while less responsible companies are the concern. He called a White House AI safety document signed by tech leaders a "good step" but short of required regulation.
The broader context here is a shift from general principles to working controls. Outside testing before release, mandatory reporting of agent breaches, and the ability to pause a system come from safety engineering and network security. Applied to large models and autonomous agents, they raise difficult questions about access for testing, monitoring systems while they run, and who has authority to order a halt.
In my view, worth noting is what Smith did not ask to slow. He is not arguing against deployment. He is arguing for a compliance layer that allows deployment to continue with independent checks. That fits Microsoft's position as both a platform operator and a large deployer, where standard testing and disclosure can lower major risks without stopping product releases.
In practical terms for teams building and buying AI, the details to watch are scope and liability. Mandatory outside approval could mean structured pre-deployment testing with set test suites and model access, or a lighter notification system. A disclosure duty for rogue agents would need clear definitions of agency, autonomy and breach, plus monitoring data to detect when an agent acted outside its limits. Kill-switches sound simple. In distributed systems with copied models, fine-tuned versions and tool connections, they are not.
Stepping back, the U.K. claim gains weight from that existing testing pipeline. If the AI Security Institute already sees frontier systems before release, turning that into a formal checkpoint is a smaller step than building a regulator from zero. The open question is enforcement. Access by agreement is different from power to block or condition a launch, especially for models built and hosted elsewhere.
Looking at the longer arc, this pattern of fast adoption is familiar. My children grew up treating smartphones and then generative tools as normal background, not novelties, and businesses have normalized them just as quickly. That speed helps explain why Smith's milk-and-washing-machine comparison connects, even if those safety models do not apply directly. Standard safety testing and clear reporting duties would make AI systems easier for companies to trust and deploy, not harder.
