AI Created a Virus That Doesn't Exist
Viral Hallucinations
An AI trained on the genetic code just did something no evolution has ever done.
It designed a working virus, not a close copy of one that already exists, but a genome with features that natural mutation would almost certainly never produce. One output lost an entire protein and stayed alive. Another added a new gene that has no counterpart in nature. A third swapped in a gene from a distantly related virus as if borrowing a spare part.
The researchers at Stanford who built the Evo models call these outputs what they are: viral hallucinations. And the biosafety question at the heart of this breakthrough is not whether the AI can do it; it already did. The question is who stops the next person from doing it with viruses that infect us.
How to Make a Virus From Scratch
The Evo 1 and Evo 2 models are large language models trained on DNA instead of text. Given enough genome sequences, they learn the grammar of life: which base pairs tend to follow which, where genes start and end, what a functional protein sequence looks like. The models are open weights and have been available on Hugging Face since early 2025.
For this experiment, the Stanford team fed Evo 1 and Evo 2 over two million additional base pairs of DNA from bacteriophages (viruses that infect bacteria) and then fine-tuned them on Microviridae, the family that includes the test virus phi X174.
Phi X174 is a tiny, well-characterized virus with 11 genes spread across roughly 5,400 base pairs. It infects E. coli. Its genome ends with a short recurring sequence, which the team used as a prompt. The right prompt length turned out to be four to nine bases: enough to signal what the model should produce, not enough to just copy back.
The raw outputs were filtered through hard constraints: the spike protein that phi X174 uses to latch onto bacteria had to be at least 60 percent identical to the real one; the genome had to be between 4,000 and 6,000 bases; no runs of the same base longer than ten; GC and AT ratios had to be reasonable. These filters narrowed the candidates from thousands to 302 plausible viral sequences.
The team was able to chemically synthesize 285 of them and insert them into E. coli.
16 Working Viruses, None of Them Normal
Most of the 285 synthesized sequences did nothing. But 16 inhibited bacterial growth, confirming they were functioning as viruses. Nine were the initial AI outputs; seven had acquired additional mutations after being inserted into bacteria.
In some ways these viruses looked exactly like their natural counterpart. The ones most similar to phi X174 (98 percent sequence similarity or above) had a 46 percent viability rate. The overall viability across all 285 outputs was just 5.6 percent.
But the interesting ones were the ones that looked nothing like the original.
One AI-designed virus had lost an entire protein and compensated with changes elsewhere in its genome. Another had an entirely new gene that no natural phi X174 carries. A third had replaced one of its own genes with a gene from a distantly related bacteriophage. Many had genes that were longer or shorter than their natural equivalents.
The researchers ran a striking analysis. In phi X174, even a single amino acid change has roughly a 20 percent chance of killing the virus. Any virus with fewer than 25 changes has only a 2.3 percent chance of being viable by random mutation. Yet nearly a quarter of the AI's outputs that had more than 25 changes were still alive, including two with over 50 amino acid alterations.
The AI did not just match random mutation. It outperformed it, consistently finding combinations of changes that kept the virus functional where evolution would have killed it.
A Cocktail That Beat Resistance
Bacteriophages have been studied as a treatment for antibiotic-resistant bacterial infections for years. The problem is that bacteria evolve resistance to individual phages fast, and cocktails of natural phages often fail against already-resistant strains.
The Stanford team tested a cocktail of the 16 AI-designed viruses against a resistant strain of E. coli and compared it to a cocktail of natural phages. The natural cocktail failed. The AI-generated cocktail overcame the resistance: the viruses swapped DNA segments and accumulated new mutations that let them infect hosts the natural phages could not touch.
This is not a ready-to-deploy therapy. Phage therapy has struggled with regulatory approval and manufacturing scale for decades. But the result suggests that AI-designed phages may be able to solve a problem (evolved bacterial resistance) that has consistently frustrated natural approaches.
The Precaution That Is Not a Wall
The Stanford team took one significant precaution during this work. They excluded all sequences from viruses that infect vertebrates from the model's training data. The thinking was straightforward: if the model has never seen a vertebrate virus, it cannot hallucinate one.
This is not the safety measure it sounds like.
The Evo models are open weights. Anyone with sufficient computing resources can download them, fine-tune them on any DNA sequence they choose, and repeat the experiment. The training data exclusion is a choice, not a technical barrier. The paper itself closes with a call for better governance of this area, including the ordering of custom DNA sequences that would be needed to synthesize any such output in a lab.
"We may want to start thinking now about preparing for the potential that someone could develop a related AI that can design a virus that targets vertebrates," the researchers write in the Science paper.
The discipline of AI biosafety has spent years worrying about language models generating instructions for bioweapons. The Evo experiment is an uncomfortable reminder that the actual risk may not come from an AI telling someone how to build a virus. It may come from an AI building the virus itself.
Sources
- Ars Technica: Large genome models used to design new viruses
- Nature: World's first AI-designed viruses a step towards AI-generated life
- New York Times: This A.I. Just Created Viruses Not Found in Nature
- Washington Post: Inside the debate over a tech breakthrough raising questions about life itself
- Engadget: AI Is Now Making New Viruses
- Science (primary paper): DOI: 10.1126/science.aec2657