10/08/2026
AI didn't simply assemble DNA from known fragments, like a building block set. It "wrote" new genomic sequences on a computer, much like developers write Python code. Scientists then synthesized 285 selected constructs in the lab. Sixteen of these proved viable: they infected E. coli bacteria, replicated within them, and destroyed the cells.
https://phys.org/news/2026-08-sixteen-ai-viruses-route-drug.html
Until now, neural networks generated text, images, and computer code. Now, the code they produce can replicate.
Why do we need artificial viruses?
It's crucial to understand that the viruses generated by AI are bacteriophages, which target bacteria, not humans. The researchers intentionally excluded human, animal, and plant viruses from the training data.
This is potentially a major breakthrough for medicine. Bacteria are becoming increasingly resistant to antibiotics, making some infections more difficult to treat with conventional drugs. Some phages created by Evo have been able to attack microbes already resistant to the natural virus, and combinations of several variants have helped slow the emergence of new resistances.
Previously, suitable phages had to be found in nature, particularly in soil, wastewater, and other environments where bacteria live. Now, it is possible to design them for a specific task.
The downside: why experts are raising the alarm
This medical advance has a downside. While generative AI can already design a complete and viable bacteriophage genome, how far can this technology go?
The current version of Evo cannot create dangerous human viruses on demand. These pathogens are much more complex, and the corresponding data were specifically excluded from the training. However, Tom Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security formulated the problem very explicitly in a commentary on the study:
The ability to create viral genomes using generative AI already exists. Yet, no governance structure is in place to regulate this technology safely.
Dangerous viruses can be tracked using their known genetic signatures. Controlling entirely new sequences will be much more difficult.
Therefore, the question is not only what AI can create, but above all, who will have this capability and for what purposes it will be used.
