AI Just Designed New Viruses From Scratch. What Could Go Wrong?

3 min readHere’s what you’ll learn when you read this story:A team of scientists led by researchers at Stanford University successfully created a new virus using AI. Much like large language models, such as ChatGPT, this AI—named Evo—uses large datasets to recognize patterns in DNA sequences. This breakthrough could have a profound impact battling antimicrobial resistant bacteria but also raises broad concerns of how AI might develop dangerous viruses or other next-gen bioweapons.“Scientists Create New Viruses Using AI” reads like the headline of a tattered newspaper, dirtied, ripped, and windswept across some post-apocalyptic landscape—a fitting epitaph for the technological hubris of the human race. But this isn’t science fiction. A new study in the journal Science details a breakthrough that could one day save millions—or lead to a new class of AI-designed bioweapons. To put it mildly: It’s complicated.The oldest battle on Earth is unseen to the naked eye. Bacteria and phages—viruses that only infect bacteria—have been locked in an evolutionary arms race for billions of years. Bacteriophages were considered our best bet against deadly bacterial infections until the early 20th century, when the arrival of antibiotics pushed phage therapy to the margins of Western medicine.But as bacteria have become wise to our medicinal techniques, they’ve become “superbugs” by developing antimicrobial resistance (AMR). Although phages are highly specific to the bacteria they attack, they offer a path for overcoming AMR, which the World Health Organization says occurs in one in six lab-confirmed cases of bacterial infection. Rapidly iterating new phages to attack certain superbugs could help keep this growing health threat at bay—and rapidly iterating is what AI does best.In a new study, a team of researchers led by scientists at Stanford University and the Arc Institute successfully used an AI model called Evo to create entirely new viruses. Evo is a genomic foundation model, a type of AI that basically treats DNA like a language so that it can invent and generate novel genetic sequences. The most recent version of the model, Evo 2, was made open-source by Arc Institute and NVIDIA in February 2025, with its weights, code, and training data available to the public. The researchers asked Evo 2, which was trained on more than two million phage genomes, to make new versions of a small, well-studied phage named Phi X-174. Working with a refined dataset that included 15,000 close relatives of the phage, Evo 2 generated thousands of candidate genomes. The team selected 302 for synthesis, successfully built and tested 285, and 16 of them were effective at infecting an E. coli bacterium.This breakthrough doesn’t mean we need to roll the credits on the human race just yet. For one, the genome of this phage contained only 5,386 base pairs. The human genome, by comparison, contains three billion. Also, the phage can only infect E. coli and nothing else, significantly constricting its ability to interact with humans. The team also worked only with non-pathogenic E. coli strains under enhanced containment. It’s also worth noting that genetic code for viruses that infect humans, animals, or plants was excluded from Evo 2’s training data specifically to prevent it from designing dangerous pathogens.“One of the most rewarding parts of this project is the creativity Evo 2 allows,” Stanford bioengineering graduate student Samuel King, lead author of the study, said in a press statement. “New doors in science are now open because of what we can do with these models.”Naturally, the technology has raised concerns that what’s behind those “new doors” could lead to AI-generated bioweapons. In 2024, for example, Chinese scientists used artificial intelligence to create novel conotoxins, which are typically found in the venom of marine cone snails. While some conotoxins have been developed into painkillers, others can be highly dangerous, and a senior U.S. government employee reportedly flagged the study as a biosecurity risk.“We’ve always made the assessment that the benefits to the world far outweigh the dangers,” University of Washington computational biophysicist David Baker, the 2024 Nobel Prize co-winner for his research into protein design, told Nature. “But, as capabilities increase, I think that’s going to be an important question to keep considering.”Like all human innovations since time immemorial, artificial intelligence is just a tool and not inherently good or evil. It’s ultimately what we do with that tool that counts.Darren lives in Portland, has a cat, and writes/edits about sci-fi and how our world works. You can find his previous stuff at Gizmodo and Paste if you look hard enough.