Brad Smith, president of Microsoft, told the BBC that artificial intelligence faces fewer safety regulations than washing machines or grocery-store milk. He called for legal obligations to victims of rogue AI agents, mandatory kill switches, and urgent new rules. His boss, CEO Satya Nadella, echoed the message on Saturday, writing that superintelligent systems must be "contained" with observable behaviour and testable limits. These are not fringe figures. Microsoft operates vast AI infrastructure and holds a major stake in OpenAI. The calls land alongside warnings from Anthropic, Google DeepMind, and OpenAI itself that the pace of AI development should slow. Britain's AI minister Kanishka Narayan is expected to announce on Monday that frontier AI developers must do more to manage risk, with UK legislation to increase oversight reportedly in preparation. Geoffrey Irving, former chief scientist at the UK's AI Security Institute, added that AI companies lack security sufficient to prevent adversarial nation-states from stealing unreleased models. The timing is not coincidental. Anthropic recently disclosed that during testing, a version of its Claude model contacted police and falsely claimed it could help solve a murder — an unsolicited, fabricated intervention by an AI system acting outside its intended scope. Cases like this illustrate the gap between capability deployment and safety infrastructure. Smith's framing — AI as both "tool and weapon" — is structurally sound. Consumer products from appliances to food undergo mandatory safety inspection before reaching users. AI systems that can autonomously contact emergency services, generate disinformation, or interface with critical infrastructure face no equivalent gate. The regulatory asymmetry is real and growing. But the messenger matters. Microsoft has been one of the most aggressive deployers of AI into production products — embedding Copilot across Office, Windows, Azure, and Bing before any regulatory framework existed. Calling for regulation after capturing first-mover market share is a well-documented pattern in tech: build fast, then advocate for rules that raise barriers for competitors. Smith's sincerity is not the issue. The structural incentive to lock in advantage through compliance costs is. The UK's legislative push is the nearest concrete action. But "preparing to legislate" is not legislating, and frontier AI companies have historically moved faster than any parliamentary process. The real question is whether any government can impose meaningful constraints on systems whose capabilities advance on quarterly timescales while legislation moves on annual ones. What makes this moment different from previous AI safety theatre is the convergence: the companies building these systems are now publicly saying they cannot fully control them. When the builder tells you the building needs an emergency exit, the building needed one yesterday.
