Researchers at Stanford and the Arc Institute trained an AI model called Evo 2 on huge libraries of DNA, the same way language models are trained on text. Then they asked it to design something that has never existed before, a complete viral genome, built from nothing.

They tested nearly 300 of the AI’s designs in the lab. Sixteen of them worked. Real, functional bacteriophages, viruses that infect and kill E. coli, made from scratch by a model that had never seen a phage in its training data. This is the first peer reviewed case of generative AI designing a functional virus genome from the ground up, published in Science.

These phages infect bacteria, not people, and some of them killed E. coli strains that had already evolved resistance to natural phages. That is genuinely useful, phage therapy against drug resistant infections is one of the most promising ideas in medicine right now, and an AI that can design new phages on demand could speed that work up by years.

But biosecurity researchers flagged something real. The screening systems that labs use to catch dangerous DNA orders were built to recognize sequences that already exist in nature. An AI that writes sequences nobody has ever seen slips right past that kind of filter. That gap needs to close, and fast, before this same approach gets pointed at something more dangerous than a bacteria killing phage.

This is what a genuinely new tool looks like in its first year. Real benefit, a real gap in the safety net, and researchers already publishing the gap in the open instead of burying it. That is closer to how this should go than not.

——

Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277

——

Sources:
https://news.stanford.edu/stories/2026/08/evo-2-ai-tool-e-coli-killer-bacteriophages
https://www.science.org/doi/10.1126/science.aec2657
https://www.techtimes.com/articles/323507/20260807/stanford-ai-wrote-viruses-no-evolution-ever-produced-biosecurity-gap-confirmed.htm

Repost this. Thanks.