AI-Made Viruses Work — and the Guardrails Are Still Catching Up

AI-Made Viruses Work — and the Guardrails Are Still Catching Up
Artificial intelligence has crossed a biological threshold: it can now help design working viruses never found in nature. The immediate targets are bacteria, not people — but the experiment has brought medical promise and biosecurity anxiety into the same frame.

The Stanford-led work, reported in Science, trained genome language models on DNA rather than ordinary text, then focused them on ΦX174, a small bacteriophage that infects E. coli. Researchers deliberately excluded viruses that target complex organisms from the models’ training data, treating bacteriophages as the safer testing ground.

After fine-tuning the systems, Evo 1 and Evo 2, on related bacterial viruses, the team generated 302 proposed genomes. They chemically synthesized 285 and inserted them into bacteria. Sixteen inhibited E. coli growth, evidence that the AI-designed sequences could function as viruses; several carried striking changes while remaining viable.

That result is not, by itself, a human-health threat. The new phages infect bacteria, and their designers see a possible route toward treatments for infections that no longer yield to antibiotics. One account called the work a source of “hope for medical advances,” while also warning that it could enable “new diseases and biological weapons.”

The sharpest dispute is over what comes next. The technical account stresses constraints: most viable designs still closely resembled ΦX174, and only 5.6% of tested outputs worked. Yet it also found that the model produced viable, heavily altered viruses far more effectively than random mutation would be expected to do.

Security specialists argue that this is precisely why oversight cannot remain focused only on modifying known pathogens. As Johns Hopkins experts Thomas Inglesby and Moritz Hanke put it, “The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”

Continue reading https://foxvector.com/stories/019fdc8b-8bb6-1e4c-73ef-26668cc8fc2c