Scientists at Stanford University and the Arc Institute have used artificial intelligence to design new functional viruses capable of infecting and killing bacteria, marking a significant advance in AI-assisted biological engineering while renewing debate around biosecurity safeguards for generative models.
The research, published in Science, involved genome language models called Evo 1 and Evo 2, which were used to generate complete genetic sequences for bacteriophages. These are viruses that infect bacteria rather than humans, animals or plants. The researchers used the well-studied bacteriophage ΦX174 as a design template and experimentally tested the AI-generated genomes.
The work represents a step beyond earlier applications of generative AI in biology that focused on individual proteins, DNA fragments or other biological components. Researchers were able to generate whole viral genomes with functional genetic architectures, demonstrating that genome-focused AI models can produce biological designs that work when synthesised and tested in laboratory conditions.
Of the hundreds of AI-generated designs examined by the research team, 16 produced viable bacteriophages. Some demonstrated stronger performance than the natural ΦX174 reference in laboratory tests. A combination of the generated phages was also able to overcome resistance in three strains of E. coli, indicating potential applications in developing treatments against bacteria that become resistant to existing therapies.
The findings could have implications for phage therapy, an area of research that uses bacteria-targeting viruses as potential treatments for bacterial infections. Interest in the field has increased alongside concerns about antimicrobial resistance and the declining effectiveness of some conventional antibiotics.
The researchers deliberately focused on bacteriophages that target bacteria. The viruses produced during the experiment were not designed to infect humans, and reporting on the research indicates that the experiment itself did not create an immediate human health threat. The distinction is important as the study has also triggered wider discussion about how similar AI capabilities should be governed.
Biosecurity experts have cautioned that advances in genome-generation technology could eventually create new risks if increasingly capable systems were applied to harmful pathogens. Researchers have consequently called for safeguards that extend beyond the AI models themselves, including screening during DNA synthesis and established biosafety controls within laboratories.
The development comes as generative AI expands into scientific research beyond text, images and software. Genome language models apply concepts similar to large language models to biological sequences, learning patterns across genetic information and using those patterns to predict or generate new sequences.
Evo, for example, was developed as a genomic foundation model capable of working across DNA, RNA and proteins. Earlier research demonstrated its ability to generate functional biological systems and produce long genomic sequences with plausible structures. The latest work extends that approach to experimentally validated whole bacteriophage genomes.
The study also highlights a growing challenge for AI governance. As models become capable of operating across scientific disciplines, policymakers and technology developers are increasingly considering how access, biological data, laboratory practices and synthesis services should be managed without restricting legitimate research.
For now, the researchers' work remains focused on bacteria-targeting viruses and potential therapeutic applications. However, the ability of AI systems to generate functional genomes introduces another dimension to discussions around responsible AI development, particularly as generative models become more capable of moving from digital outputs to designs that can be physically created and experimentally tested.