Something shifted in Westminster last week. AI moved from a background policy concern to an active security alarm, with cabinet discussions, MI5 warnings, and parliamentary hearings converging in a single seven-day stretch. The catalyst was not a single event but an accumulation: AI minister Kanishka Narayan warned that state-level hacking was "likely, if not inevitable," MI5 chief Ken McCallum called AI a "supercharger" for cyberattacks and bioweapon threats, and Liam Byrne's business committee grilled executives from Anthropic, OpenAI, DeepMind, and Meta on whether model testing should be mandatory. The political backdrop matters. Multiple sources credit the shift from Keir Starmer to Andy Burnham's leadership with creating distance from Big Tech. Byrne was blunt: "The Starmer administration was much too close to the big tech companies. Burnham is not." Another senior figure described the previous posture as having "drunk the tech-bro Kool-Aid." Whether this represents genuine strategic recalibration or a new government establishing its identity by contrast is an open question. The substantive worry has two layers. The immediate concern is cybersecurity and bioweapons — McCallum flagged extremist groups using AI to design biological agents, though experts note a gap between designing pathogens and weaponising them. The deeper concern, articulated privately by a cabinet minister, is the concentration of superintelligent capability in the hands of bad actors — not killer robots, but unchecked power asymmetry. Open-source AI models are the specific vector that alarms the security establishment. Geoffrey Irving, former chief scientist at the AI Safety Institute, told the Guardian that open-source models are "very likely already powerful enough to do serious damage" and typically catch up to proprietary models within a year. Safety guardrails on these models "either don't exist or can be trivially removed." This is the core policy problem: the most dangerous capabilities are the ones least amenable to regulation. Yet the government is simultaneously fighting itself. Defence sources flagged the cancellation of the AI programme at the Defence Science and Technology Laboratory at Porton Down as shortsighted. The MoD countered with £400m in AI-focused investment over four years, arguing AI should be "integrated across all workstreams, rather than a standalone programme." Former cabinet secretary Gus O'Donnell warned that ministers needed to weigh "the risk of being left behind" as heavily as the need for guardrails. The international dimension adds urgency and constraint. Trump's administration reportedly asked AI companies not to share latest models with the UK's AI Safety Institute, though Narayan says cooperation has been retained. Byrne sees the upcoming US midterm elections as a window — if Democrats retake Congress, international regulatory momentum could shift. The UK's G20 presidency is being positioned as the vehicle for collective action. What emerges is a government caught in a genuine policy trilemma: regulate hard enough to address real safety risks, adopt fast enough to capture productivity gains in public services, and maintain access to US-controlled frontier models despite a hostile White House. Solving any two of these likely makes the third harder. The week of alarm bells was real, but alarm bells are not a strategy.
