AI genome models have now generated functional viral DNA, raising biosecurity questions

Advances in AI-driven biology have historically centered on protein design, where machine learning can directly reshape the chemistry and machinery of life. But a new frontier has emerged: using large genome models—neural networks trained on massive DNA datasets—to generate entirely novel viral sequences.

Researchers at Stanford University have now demonstrated that these models can create functional viral genomes that infect bacteria. While all viruses produced so far remain variants of existing organisms, they exhibit features that would be difficult for nature to evolve on its own, marking a significant shift in what's possible with machine-generated genetic code.

How genome models work#

Large genome models apply the same predictive logic as language models, but to DNA's four-letter alphabet (A, T, C, and G). Unlike human text, however, genomes contain regions where sequence order is critical for function—promoters, genes, regulatory elements—interspersed with stretches where variation carries no consequence. These models must learn to recognize biological context at a level humans often don't fully understand.

When fed enough genomic data, they develop surprising intuition. Bacterial genes with shared functions tend to cluster together in genomes. Prompt the model with part of such a sequence, and it generates related DNA that encodes proteins with matching biochemistry—sometimes producing working sequences unlike anything discovered in nature.

The viral leap#

Prior breakthroughs showed that genome models could output DNA encoding functional bacterial proteins and replicate gene structures found in complex organisms. The new work extends this to complete viral genomes: self-contained packages that can infect and replicate within cells.

The Stanford team emphasizes that the viruses they created are closely related to known species. Yet they carry distinct features that evolution wouldn't easily produce—suggesting the models are not merely interpolating between existing examples but generating genuinely novel sequences.

Security ahead#

These results raise pressing concerns about dual-use risks. The current work focuses on bacteriophages—viruses that target single-celled organisms. But the implications are clear: if similar models could be adapted to design pathogens that infect vertebrates, the consequences could be severe.

The Stanford researchers argue the time to address this risk is now, before the capability exists, rather than waiting for a crisis. Whether through technical safeguards, regulatory oversight, or responsible disclosure norms, the field may need guardrails to ensure genome models remain tools for medicine rather than instruments of biological harm.