An AI just designed viruses that have never existed anywhere in nature. And the screening tools we use to catch dangerous DNA literally can't see them.
What Actually Happened
Stanford University and the Arc Institute published findings in Science showing that Evo 2, a genome-design AI trained on DNA the way language models are trained on text, designed complete viral genomes from scratch. The results:
16 of those genomes proved fully functional when synthesized and introduced to bacteria -- real bacteriophages that infected and killed E. coli in a Palo Alto lab.
Because these genomes have no evolutionary history, they match nothing in the databases labs use to screen synthetic DNA orders for known dangerous sequences.
Johns Hopkins Center for Health Security reviewed the paper and described it as raising "urgent biosafety and biosecurity questions," pressing for a legal requirement that DNA synthesis companies screen both sequences and customers -- a requirement that doesn't currently exist in U.S. law.
The success rate isn't perfect: only about 5% of the AI's candidate genomes actually worked. So we're not at design-a-virus-on-demand. Yet.
The Real Story Isn't the Danger Level Today
The viruses here only target bacteria, not humans, and Evo 2 was deliberately trained without human, animal, or plant pathogen genomes. But the underlying gap is the real story: screening infrastructure built to catch known threats has no way to catch something genuinely new. As researchers put it, the ability to do this now exists -- the governance to safely steer it doesn't.