arXiv:2609.16213 maps AI-driven digital biosecurity uplift against persistent wet-lab barriers

The paper catalogs threat pathways from information retrieval through design, procurement, synthesis, testing, and release. It distinguishes digital assistance, where large language models and biological foundation models deliver measurable uplift, from physical stages where automated labs close only partial segments of the design-build-test-learn loop. Evidence cited includes benchmark results on sequence generation and evasion alongside wet-lab experiments showing no equivalent acceleration in execution accuracy.

Alignment methods developed for general models fail to transfer because biological foundation models optimize over sequence spaces with direct functional outputs. Interpretability techniques are positioned as the primary audit mechanism for verifying removal of hazardous capabilities rather than post-hoc refusal training. The review therefore rejects single-point controls in favor of defense-in-depth thresholds tied to model access, laboratory capability, and material screening.

Operational implications follow directly from the data. Organizations running automated laboratories must implement staged access linked to capability evaluations, while synthesis providers require sequence screening calibrated to current model performance. Policymakers gain a concrete mapping from observed benchmark deltas to proportionate obligations across the ecosystem.

Next steps center on empirical validation of the proposed thresholds through red-team exercises that combine current models with partial automation, scheduled within 18 months of publication.