Anticipating Unimagined Biothreats in the Age of AI

Current biosecurity defenses against AI-enabled biological risk are powerful and already producing measurable gains, yet they remain organized around a single, observable object of concern: a sequence, an organism, a toxin, or a capability already judged dangerous. Biological harm can also arise from relationships among components that, taken one by one, appear ordinary. Through three thought experiments---an altered vector, an uncharacterized animal virus steered toward human biology, and a bacteriophage paired with a compatible bacterial host---we show that danger may become visible only when several ordinary actions are considered together. Fragmentation of information across providers can hide that relational risk even when every individual screen functions correctly. We therefore argue that the unit of oversight must expand from the isolated transaction to the project itself for closed tools. Prevention has limits that no architecture can erase; resilience remains indispensable. The aim is not alarm but vigilance: biosecurity is not finished work.