Stanford AI Creates 16 Novel Viruses From 9 Trillion DNA Letters

3 articles · Updated · Tom's Hardware · Aug 8

Of 285 AI-designed viral genomes synthesized in the lab, 16 assembled into functioning phages that infected E. coli and reproduced, according to a Science study by Stanford and the Arc Institute.
After training Evo on more than 9 trillion nucleotides across 128,000 genetic sequences, researchers fine-tuned it on the 5,386-nucleotide Phi X-174 virus and 15,000 close relatives, then narrowed 700,000 candidate genomes to 285.
The resulting viruses were close variants of Phi X-174 rather than radically different organisms, and the team said it excluded human-pathogen data and produced viruses harmless to humans.
In a resistance test, cocktails of the AI-designed phages evolved to overcome three E. coli strains that had become resistant to natural Phi X-174, pointing to possible antimicrobial uses.
The work also sharpens biosecurity concerns because it shows generative AI can derive viable biological designs from DNA patterns, even as guardrails and regulation remain limited.

Sources

Left67%

Center33%

Tom's Hardware12d ago

AI Model Creates 16 Novel Viruses After Learning DNA Patterns from 9 Trillion Nucleotides

The New York Times14d ago

Arc Institute AI Creates Novel Viruses Not Found in Nature

Ars Technica14d ago

Stanford Researchers Use Large Genome Models to Design New Bacteria-Infecting Viruses

14 Sources

Could AI-designed synthetic viruses become the ultimate cure for the world's escalating antibiotic resistance crisis?

If artificial intelligence can invent entirely new viruses in a lab, what stops this technology from engineering the next global pandemic?

The 2025 Breakthrough in AI-Generated Genomes: How Evo and StripedHyena Models Are Redefining Synthetic Biology, Medicine, and Biosecurity

Overview

In 2025, Stanford and Arc Institute researchers used the Evo genomic language model to design 302 complete viral genomes, leading to 16 functional bacteriophages that could overcome bacterial resistance. This breakthrough was made possible by new AI architectures like StripedHyena, which expanded the model’s ability to process much longer DNA sequences. As Evo and similar models were released as open source, users quickly bypassed built-in biosecurity safeguards, raising concerns about dual-use risks. In response, U.S. policymakers began moving toward centralized oversight, introducing new legislation to address regulatory gaps and ensure responsible governance of AI-designed life.

...