From 9 Trillion Nucleotides to a New Virus: The Evo Model’s Breakthrough

Researchers at Stanford University and the Arc Institute in California have used artificial intelligence to design and produce entirely new viruses that do not exist in nature, according to the New York Times. The team trained a large AI model named Evo — whose underlying architecture is comparable in some ways to ChatGPT, developed by OpenAI — on approximately nine trillion nucleotides drawn from the genomes of animals, plants, microorganisms and existing viruses.

Once trained, Evo was able to generate novel DNA sequences that represent complete virus recipes. The scientists synthesized these designer DNA molecules and inserted them into bacteria. The modified bacteria then began producing the AI-designed viruses, which subsequently infected and replicated within other bacteria, confirming that the synthetic viruses were viable and functional.

The viruses were bacteriophages — viruses that infect bacteria, not humans. Nonetheless, the demonstration marks the first known case of an AI model autonomously generating workable viruses from scratch, crossing a threshold that blends synthetic biology with generative AI in a way that has both biomedical promise and significant risk.

The Promise and Peril of AI-Designed Viruses

Evo’s Novel Viral Designs: A Scientific Milestone

This experiment represents a leap in AI’s creative capability in the life sciences. Where earlier models predicted protein structures or optimized known genetic sequences, Evo can compose entirely new organisms. The underlying model, trained on the raw language of DNA across all kingdoms of life, appears to have internalized the combinatorial rules that make a virus functional. For medicine, this opens the door to designing custom bacteriophages to combat antibiotic-resistant infections, creating synthetic vectors for gene therapy, or accelerating vaccine development by generating harmless viral shells.

Advertisement

The Dual-Use Dilemma Intensifies

No matter the benign intent of the study, placing virus design into the hands of an AI that can iterate far faster than human researchers raises profound biosecurity questions. The method used to create bacterial viruses could, in principle, be adapted to mammalian systems — and eventually to human pathogens. Security experts have long warned that generative AI applied to biology could lower the barrier to weaponizing pathogens; this research makes that concern tangible rather than hypothetical.

The fact that the training data included genetic material from many species — including viruses that cause disease — means the model has been exposed to the DNA patterns of hazardous agents. While the published work focused exclusively on bacteria-infecting viruses, the same pipeline could be repurposed by a malicious actor who gains access to a comparable trained model and the necessary wet-lab equipment. The study’s publication in itself may prompt calls for tighter oversight in the vein of the gain-of-function research debates.

What the Dual-Use Breakthrough Means for Labs, Industry, and Oversight Bodies

  • Research institutions working with AI-driven synthetic biology should immediately implement mandatory dual-use review protocols before generating novel organisms, even for proof-of-concept studies like this one.
  • Biotech and pharma companies that use generative AI models for drug or therapy design must screen all AI-generated sequences against known pathogen databases to prevent accidental production of viable infectious agents.
  • International governance bodies such as the WHO and the Biological Weapons Convention review conferences should begin updating their guidance to explicitly address AI-generated biological agents, treating them as a new category of dual-use technology.
  • Funding agencies should require that grants involving generative AI in biology include a biosecurity impact assessment, and may consider restricting the release of fully trained models that can generate full pathogen genomes.