Researchers at the Arc Institute in Palo Alto and Stanford University have used artificial intelligence to design new viruses that can infect and multiply inside bacteria. A study published in the journal Science on August 6, 2026 explains how the team used AI to generate about 700,000 possible viral genetic codes, picked the 285 most promising ones, created their DNA in a lab, and inserted them into bacteria. Of those, 16 turned into working viruses. Some reproduced even faster than a well-known reference virus called Phi X-174 (Engadget).

The AI models, called Evo 1 and Evo 2, work similarly to ChatGPT. But instead of learning the patterns of human language, they learned the patterns of DNA, which is written in chemical letters called nucleotides. By studying trillions of these letters, the models learned the “grammar” of genetic code well enough to write new, complete viral genomes. For this experiment, the team trained the models on about 15,000 viruses related to Phi X-174, a small virus that only infects E. coli bacteria (Engadget). The Stanford team described their tool as writing whole genomes to create bacteria-fighting viruses (Stanford News).

The researchers deliberately designed the experiment so none of the viruses could infect humans, animals, plants, or fungi (Engadget). In lab tests, a mixture of the AI-designed viruses killed E. coli bacteria that had resisted natural viruses meant to fight them (The Guardian). The Science paper confirms that the lab-made viruses were functional: they copied themselves inside bacteria and produced working offspring (Science).

The medical purpose is straightforward. Viruses that attack bacteria, called phages, have been studied for decades as a possible alternative to antibiotics, especially for infections that no longer respond to existing drugs (University of Reading). As antibiotic resistance grows worse worldwide, the ability to use AI to design phages aimed at specific drug-resistant bacteria would meet a real and urgent need.

This work builds on several recent advances in using AI for biology. AI tools have already discovered roughly 70,000 new viruses by sifting through genetic material from extreme environments like salt lakes and deep-sea vents (Nature). A system called AlphaFold has mapped how virus proteins are shaped, helping scientists understand relationships between viruses that cause diseases like dengue and hepatitis C (Nature). AI is also being used to predict how viruses might evolve and to help plan responses to outbreaks (MDPI). The Evo study takes these capabilities a step further: not just finding or studying viruses, but creating functional ones from scratch.

The possibility of misuse is where the conversation gets harder. To design viruses that could harm people, AI models would need to accurately predict what makes a virus dangerous, including how it is built, how sick it makes you, and how easily it spreads (NCBI.) The viruses in this experiment are nowhere near that level. They target bacteria, not human or animal cells. But the overall process, from AI-generated genetic code to a working virus, could in principle be applied more broadly. Several pathways exist through which AI advances could enable the deliberate release of harmful biological agents (Safe.ai). In October 2025, RAND reported that concerns about AI-enabled pathogen design are growing, though the exact risks and timelines are still uncertain (RAND). By May 2026, scientists were actively debating whether to limit biological AI software to reduce threats from AI-designed viruses, toxins, and other weapons (Nature).

This tension is not new. Every major biological tool, from the ability to cut and paste DNA in the 1970s to the gene-editing technique CRISPR more recently, has raised the same question: how do you balance open scientific progress with the risk that someone could misuse the technology? In 1975, scientists held a conference at Asilomar and agreed on voluntary safety rules for DNA research that lasted for decades. After a researcher in China used CRISPR to edit human embryos in 2018, the gene-editing community developed its own guidelines. The question now is whether similar rules can be created quickly enough for AI models that are improving fast.

What sets genome language models apart from earlier biotechnology breakthroughs is speed and access. A lab using traditional methods might need months to engineer a single new virus. Evo produced hundreds of thousands of candidates computationally in a fraction of that time, and the slowest part became physically creating and testing the DNA. As the cost of making DNA keeps falling and the models keep improving, that bottleneck gets smaller.

The practical benefit is clear. Phage therapy has struggled in part because finding the right natural virus to kill a particular bacterium is slow and hit-or-miss. An AI model that can produce working, targeted viruses on demand would change that. The 16 functional viruses from this experiment, some outperforming the natural reference virus, show the approach works at an early stage. Turning it into a medical treatment approved for human use will require much more testing, including safety studies, understanding how the immune system reacts, and developing regulatory rules.

In my view, the governance gap may be the more urgent problem. The technology to create working viral genomes now exists in a published, peer-reviewed form. The safety frameworks to control how it spreads do not. The debate among scientists over software limits, reported in Nature this past May, shows that people are aware of the gap. Whether that awareness leads to effective rules before the next advance shrinks the distance between this experiment and something more dangerous is the question that should matter most to anyone paying attention.