MI5 chief Sir Ken McCallum has flagged AI as a potential game-changer for bioweapons development, and the evidence trail he's pointing at is real. Anthropic reported five cases of scientists in unnamed countries attempting to use Claude for potentially dangerous work on viruses and toxins, including one effort to introduce mutations in highly pathogenic bird flu to enhance mammalian transmission. An MIT exercise demonstrated that non-science students could, within an hour, extract from AI chatbots the names of pandemic pathogens, synthesis methods, and under-screened DNA suppliers. But the experts interviewed here draw a clear line between information access and operational capability. Dr Filippa Lentzos at King's College London frames the real threat tier correctly: AI doesn't remove the need for expertise, it amplifies expertise already present. The most immediate danger sits with well-resourced nation states and organizations that already have skilled scientists, high-containment labs, and full equipment suites. For them, AI is an accelerant on an existing capability stack. The garage-lab terrorist scenario gets the headlines but faces compounding physical barriers. Prof Martin Pačesa at the University of Zurich offers the sharpest analogy: any graduate physicist can sketch a better nuclear bomb, but where does he get the uranium? Sonia Ben Ouagrham-Gormley at George Mason identifies tacit knowledge — the unwritten experimental know-how that historical state bioweapons programs depended on — as a bottleneck AI is unlikely to fill. A recipe is not a capability. The dual-use problem makes governance genuinely difficult. Los Alamos scientists trained an AI to predict how marine snail neurotoxins bind to human cells — legitimate antidote research. Chinese researchers then built on it to design new toxins, claiming drug-development intent. In April, scientists created the first AI-designed viruses (bacteria-targeting phages with medical potential), immediately raising biosecurity alarms. The line between defensive research and weapons development runs through intent, not through the science itself. The governance gap is the real story. Prof Tom Inglesby at Johns Hopkins confirms: no country has formal policies preventing AI-assisted bioweapons development. No mandatory pre-release evaluations exist for AI models. DNA synthesis screening remains largely voluntary, and AI-designed novel genetic sequences can sidestep existing screens entirely. Anthropic disclosed its cases; the majority of AI companies are not announcing what, if anything, they do. The Indian ricin plot — a doctor and associates using AI chatbot queries and AI-driven search to attempt ricin production with kitchen tools — illustrates the current low end: crude, probably ineffective, but indicative of intent meeting newly accessible information. The high end is quieter and more dangerous: state-backed researchers using AI to optimize pathogen transmissibility or drug resistance within fully equipped biosafety labs, where the work is indistinguishable from legitimate pandemic-preparedness research until it isn't. The 20-year trajectory is clear. AI capability grows monotonically. Physical barriers erode more slowly but do erode — synthetic biology costs drop, DNA synthesis becomes more distributed, screening remains voluntary. The window for establishing governance infrastructure is now, before the intersection of AI capability and biological accessibility moves from state-level concern to non-state reality.
