AI Designs Functional Viral Genomes From Scratch, Pushing Biosecurity Oversight to Its Limits

Researchers at Stanford University and the Arc Institute have demonstrated, for the first time, that artificial intelligence can generate complete, functional viral genomes without any human-guided gene editing. The work, published August 6 in the journal Science, is a proof-of-concept that the field of generative genomics has cleared a threshold that biosecurity analysts have long flagged as a policy trigger.

The team, led by assistant professor Brian Hie and graduate student Samuel King, used two genome language models, Evo 1 and Evo 2, to generate complete phage genomes. These systems don't work like chatbots. According to reporting by BetaNews reviewing the study, they operate on genetic sequence data rather than written text. The target was ΦX174, a small, well-characterized bacteriophage that infects E. coli. It's a tractable test case precisely because its biology is so well understood, which makes anomalies easier to catch.

The numbers matter. According to Xinhua's coverage of the paper, 285 of the AI-generated designs were successfully synthesized and assembled in E. coli cells, and 16 of those produced viable bacteriophages that reproduced and killed bacteria in laboratory tests. Some of the designed phages outperformed the natural phiX174 in direct competition. A mixture was also effective against E. coli strains that had already evolved resistance to the natural phage, which the authors noted points toward potential phage therapy applications.

Those are the legitimate civilian applications. The dual-use framing is unavoidable here, and the researchers didn't avoid it. As News-Medical reviewed from the paper, King et al. emphasize that the ability to design and synthesize functional AI-generated genomes "introduces important biosafety and biosecurity concerns and underscores the need for expert oversight and robust safeguards throughout the design process." They argue that model-level protections, such as excluding sensitive viral sequences from training data, may provide an additional layer of risk mitigation. Evo 2, per the study, was developed with one such restriction: its developers excluded viruses that infect animals, plants, and humans from its training data. That's worth noting, but it's a vendor-imposed control, not a regulatory requirement, and the virus-host specificity of a language model can't be treated as a hard guarantee.

The accompanying Perspective in Science, authored by Thomas Inglesby and Moritz Hanke of Johns Hopkins University, is where the policy ask becomes explicit. As reported by Xinhua from that Perspective, the two researchers called for legally required screening of synthetic genetic-material orders for potentially dangerous sequences, rather than relying on voluntary safeguards. That demand has circulated in biosecurity circles for years. The novelty is that a demonstration-grade paper, not a theoretical argument, is now behind it.

The jurisdictional problem underneath all this is real and poorly resolved. A separate analysis published in Science on June 5 by Professor Seokbeom Kwon of KAIST, reviewed by ETL from the EurekAlert release, examined roughly 600,000 research papers and found that the share of dual-use research directly involving the U.S. federal government has declined over time, while the share involving foreign funding or non-U.S. institutional authorship climbed from 35 percent to 54 percent over the period studied. Kwon's finding, published as "Dual-use research under scrutiny" (DOI: 10.1126/science.aee2479), is not that oversight is wrong. It's that unilateral domestic tightening imposes disproportionate costs on U.S. science while comparable research continues outside U.S. jurisdiction. Executive Order 14292, signed in May 2025, intensified federal scrutiny of gain-of-function research, but that order's reach stops at the U.S. funding boundary.

The confidence assessment on the biosecurity risk from this specific paper is low to moderate. The phages designed here target E. coli. The jump from bacteriophage design to human-pathogen design isn't trivial. But the capability demonstration is real, and the tools that enabled it are not confined to Hie's lab. The question the Johns Hopkins Perspective poses, and that no regulatory body has answered, is whether the screening infrastructure for synthetic DNA orders can keep pace with generative genomics as it scales. Right now, the honest answer is that it hasn't been tested at scale under legally binding conditions.

Sources cited:
- Science (King et al., Aug. 6, 2026) (https://www.science.org/doi/10.1126/science.aee2479)
- BetaNews (https://betanews.com/article/ai-designed-viruses-stanford-arc-institute/)
- News-Medical (https://www.news-medical.net/news/20260807/Researchers-use-AI-to-design-functional-bacteriophage-genomes-from-scratch.aspx)
- Xinhua (https://english.news.cn/northamerica/20260807/e5e9280d2df74c31a40e2b0aae00d226/c.html)
- EurekAlert (KAIST / Kwon, June 5, 2026) (https://www.eurekalert.org/news-releases/1131017)
- Science (Kwon, "Dual-use research under scrutiny") (https://www.science.org/doi/10.1126/science.aee2479)