Researchers at Stanford University have demonstrated that large genome models, specifically Evo 1 and Evo 2, can successfully design functional, synthetic viruses. By training these models on vast DNA datasets and fine-tuning them with sequences from the Microviridae family, the team enabled the AI to generate novel genomes based on the ΦX174 bacteriophage. Using prompts of four to nine bases from the ΦX174 start sequence, the models produced diverse outputs ranging from near-exact copies to highly divergent sequences. To filter results, researchers applied strict constraints, such as requiring a spike protein with at least 60% identity to the original and maintaining a genome length between 4,000 and 6,000 bases. Out of 285 chemically synthesized sequences, 16 were viable in inhibiting E. coli growth. Notably, while random mutations with over 25 amino acid changes typically result in non-viable viruses, nearly 25% of the AI-generated sequences with more than 25 changes remained functional. These AI-designed viruses also proved effective in overcoming bacterial resistance, outperforming natural bacteriophage cocktails in infection trials. Despite the focus on bacteriophages, the researchers warn that these models could potentially be adapted to design viruses targeting vertebrates, necessitating improved governance for custom DNA synthesis.