AI-Designed Viruses Spark Biosecurity Alarm After Successful Lab Synthesis

Researchers from the Arc Institute and Stanford University have successfully used generative AI to design and synthesize entirely new viruses that never existed in nature. The experiment demonstrates that genomic language models can now compose functional biological entities from scratch, a breakthrough that offers potential medical utility while raising urgent biosecurity warnings.

Using a genomic AI model called Evo, the team generated approximately 700,000 potential viral genomes. To achieve this, Evo was trained on a massive dataset of genetic sequences found in nature, totaling over 9 trillion nucleotides. From the initial pool of designs, researchers synthesized about 300 candidates in a laboratory setting. This process resulted in 16 functional novel viruses, known as bacteriophages, which successfully targeted E. coli bacteria. The findings were published in the journal Science in August 2026.

The Genomic Frontier

The research was designed to test whether AI could move beyond analyzing existing DNA to actually composing new, viable genetic code. To mitigate immediate risks during the study, the researchers intentionally avoided training the model on viruses that impact humans. The resulting bacteriophages are specialized to attack bacteria, meaning they pose no direct threat to human health. However, the ability to move from a digital AI design to a physical, functioning virus marks a significant shift in synthetic biology.

Implications for Biosecurity

While the ability to create custom phages provides a promising path for treating antibiotic-resistant "superbugs," the experiment has ignited a fierce debate over the lack of guardrails for generative AI in genomics. Experts argue that the technical capability to design lethal or highly transmissible pathogens now exists, yet the regulatory frameworks to prevent such misuse are nonexistent. Dr. Moritz Hanke and Tom Inglesby noted that while the ability to compose viral genomes using generative AI now exists, the governance to safely steer it does not.

Divergent Perspectives

Not all experts view the risk as immediate or primary. Tom Ellis, a Professor of Synthetic Genome Engineering at Imperial College London, suggested that the threat of full AI-driven genome design may be overblown. Ellis argued that the more likely and immediate pathogenic threat comes from taking existing pathogens and applying gain-of-function changes to their genomes, which is a simpler process than designing a virus from scratch.

What's Next

As genomic AI continues to evolve, the scientific community is now grappling with how to implement safety layers for models like Evo. The primary challenge remains the creation of international governance standards that can keep pace with the speed of AI development. Future scrutiny will likely focus on whether synthesis providers can effectively screen AI-generated sequences to prevent the creation of harmful biological agents.