How much more infectious were the AI-designed viruses?

The top-performing AI-designed phage, Evo-Φ69, showed an expansion rate between 16 fold and 65 fold over a six-hour infection period, compared with the natural ΦX174's expansion rate of 1.3 fold to fourfold over the same period. Across three experiments, ΦX174 never ranked higher than third most effective at killing E. coli bacteria, and did not even place in the top five most infectious bacteriophages in one test.

What is the potential medical use?

The bacteriophages could be used for phage therapy to target multidrug-resistant bacterial infections. The team demonstrated this by evolving three E. coli strains resistant to ΦX174, and found that mutations in the artificial bacteriophages were able to overcome bacterial resistance through recombination of multiple synthetic phage strains. The team is now working toward clinical applications for both human bacterial infections and those targeting crops.

How much work did the AI models require to design functional viruses?

Neither Evo model could simply produce a functional virus genome without substantial help. Hie's team introduced the models to additional genetic samples, used engineered prompts with ΦX174-specific sequences, and applied inference-time guidance to refine outputs. This process generated 302 candidate genomes, of which 285 produced full genomes and 16 inhibited E. coli growth.

Large Language ModelsAI Safety & AlignmentOpen-Source AIHacker NewsAug 6, 2026Aug 7, 2026, 06:00 JST

Stanford AI designs viruses far deadlier than nature—in working lab test

Send on LINE

3 Key Points

What happened
Stanford bioengineers used AI models Evo 1 and Evo 2 to design synthetic bacteriophage (virus) genomes that, when produced in the lab, proved far more infectious than the natural ΦX174 bacteriophage they were based on. The top performer, Evo-Φ69, showed an expansion rate between 16 fold and 65 fold over six hours, compared with ΦX174's 1.3 fold to fourfold expansion. This marks the first time an AI-generated genome has been produced in the real world and found functional in tests, according to team lead Brian Hie.
Why it matters
The work demonstrates that generative AI can design biological systems more effective than nature—in this case, phages that kill antibiotic-resistant bacteria, potentially useful for phage therapy. However, the same capability could theoretically be applied to human viruses; the HIV genome is only about 10,000 bases and the coronavirus genome about 30,000 bases, similar in scale to bacteriophages. Hie argues the barrier to creating dangerous human viruses remains high because his team required extensive computational work and custom training data—but the feasibility concern has alarmed the biosecurity community.
What to watch
The paper is now under peer review. Hie's team is working toward clinical applications for both human bacterial infections and crop pathogens. The open-source Evo models that enabled this work are publicly available, raising questions about what actors with more resources might achieve.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The Stanford team's achievement rests on a foundation of prior work: the bacteriophage ΦX174 has been extensively studied and genetically mapped since the early 20th century, providing a well-understood starting point. By feeding the Evo models this natural template along with custom prompts and additional genome training data, the researchers were able to guide the AI toward variations with beneficial mutations. The resulting synthetic genomes are not complete viruses on their own—they are cell-hijacking genetic codes that require researchers to inject them into E. coli cells to activate. Once injected, they replicate naturally, a process Hie notes is straightforward to scale in the lab.

The practical barrier to creating dangerous human viruses remains substantial, according to Hie. The Evo models were trained on bacteriophage data, not human virus genomes, so adapting them to human pathogens would require the kind of extensive computational and experimental legwork his team performed—compounded by the absence of human virus pretraining data. Hie contends this makes designing a bioweapon with AI harder than simply engineering viruses from natural sources. Yet the biosecurity concern is not unfounded: the HIV genome at 10,000 bases and the coronavirus genome at 30,000 bases fall within the scale the team demonstrated it can work with, and the open-source Evo models are publicly available.

FAQ

How much more infectious were the AI-designed viruses?: The top-performing AI-designed phage, Evo-Φ69, showed an expansion rate between 16 fold and 65 fold over a six-hour infection period, compared with the natural ΦX174's expansion rate of 1.3 fold to fourfold over the same period. Across three experiments, ΦX174 never ranked higher than third most effective at killing E. coli bacteria, and did not even place in the top five most infectious bacteriophages in one test.
What is the potential medical use?: The bacteriophages could be used for phage therapy to target multidrug-resistant bacterial infections. The team demonstrated this by evolving three E. coli strains resistant to ΦX174, and found that mutations in the artificial bacteriophages were able to overcome bacterial resistance through recombination of multiple synthetic phage strains. The team is now working toward clinical applications for both human bacterial infections and those targeting crops.
How much work did the AI models require to design functional viruses?: Neither Evo model could simply produce a functional virus genome without substantial help. Hie's team introduced the models to additional genetic samples, used engineered prompts with ΦX174-specific sequences, and applied inference-time guidance to refine outputs. This process generated 302 candidate genomes, of which 285 produced full genomes and 16 inhibited E. coli growth.

Hacker NewsRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.