Figure 2： A biological weapon risk chain (Brady & Lee et al., 2026), with steps highlighted from Nelson and Rose (2023).

risks. AI agents demonstrate increasing autonomy across research functions, while a parallel design, genome prediction, and other complex tasks.4 While AI agents - and their interaction with tools including BTs - hold promise for advancing public health and societal welfare, these same capabilities introduce potentially serious risks, ranging from unintended consequences to (BW) development (Drexel and Withers 2024, Bengio et al. 2025, Chaves de Lima et al. 2024). Historically, a lack of sufficient expertise or effective application of that expertise towards creating BWs has been a barrier for threat actors achieving BW aims (Ouagrham-Gormley 2014. Washington et al. 2024). Emerging AI technologies may erode these constraints (Bengio et al. 2025): Luckey et al. (2025), for example, highlight the propensity for AI to enable a wider range of actors across the sophistication spectrum to conceptualize, plan, and potentially conduct BW Nelson and Rose (2023). Although depicted sequentially, real-world attack planning often unfolds plans. At present, our team's evaluations (e.g. Brady & Lee et al., 2026) adopt a primarily 4 Webster et al. (2025) identify over 1000 AI-enabled biological tools, including several dozen state of the art tools that map to the BW risk chain. See also Del Castello and Willis 2025, Moulange et al. 2024, Peppin et al. 2025 and