AI Has Learned to Build Viruses Before Regulators Learn to Stop Them

AI Has Learned to Build Viruses Before Regulators Learn to Stop Them
Artificial intelligence has crossed another biological threshold: it can now propose virus genomes that work in the lab. The achievement points toward new weapons against drug-resistant bacteria—and toward a governance gap scientists say is widening just as fast.

The work began with large genome models, which apply the next-token logic of language AI to DNA’s four chemical letters. Stanford-led researchers trained Evo models on bacteriophage sequences, deliberately excluding viruses that infect complex organisms, then tuned them to the family containing the E. coli virus ΦX174.

Using short snippets of ΦX174 as prompts, the team generated 302 candidate genomes. They synthesized 285 and inserted them into bacteria; 16 inhibited E. coli growth, evidence that they functioned as viruses. Most viable designs still resembled their natural template, underscoring how unforgiving viral genomes can be. Yet several carried striking changes, including altered gene lengths, a swapped-in gene from a distant virus, and in one case the loss of a viral protein offset elsewhere in the genome.

That is the hopeful reading: AI may help engineer phages able to overcome bacterial resistance, a potential answer to infections antibiotics can no longer reliably treat. But the same result has sharpened the darker interpretation. The authors of the security commentary put it bluntly: “The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”

Existing safeguards were built largely around modifying known pathogens, critics argue, while generative genomics can produce designs not found in nature. The Stanford team’s choice of bacteria-only viruses and extensive screening shows caution can be built into an experiment. It does not settle who sets the rules when the models become more capable—or when less cautious users gain access to them.

Continue reading https://foxvector.com/stories/019fd951-fe6e-10cf-719c-1a8db39e477f