Artificial-intelligence systems have been displaying some eye-opening new abilities in recent weeks: breaking out of their enclosures, hacking other companies, lying to people.
And now, for the first time, an AI model has created a new virus. A whole family of them, in fact.
In a study published in the journal Science, researchers at Stanford University and elsewhere said they successfully directed an AI model to create a collection of simple viruses that only infect bacteria and pose no threat to humans.
In an interview, two of the authors said their research could help develop designer viruses that could be used to eliminate drug-resistant bacteria, which according to the World Health Organization kill millions of people every year.
“If we develop responsibly, I think it can lead to tremendous benefits in human health,” said Brian Hie, a computational biologist at Stanford and an author on the study.
The study —which documents experiments that concluded about a year ago—arrives at a moment of heightened anxiety regarding danger posed by increasingly powerful AI models. Starting in mid-July, OpenAI, Anthropic and Meta Platforms disclosed incidents in which their models escaped contained environments, got onto the internet and hacked other companies’ computers. Anthropic’s Mythos practiced deception, while OpenAI models swapped tips on a secret message board.
The potential use of chatbots or agents to develop deadly pathogens has long been a concern among AI safety experts. Last year, after the release of an upgraded version of ChatGPT, hundreds of people queried the chatbot to ask it how to produce biological weapons and poisons, and it responded by providing instructions that biology and terrorism experts judged as accurate, The Wall Street Journal previously reported.
Experts said the Stanford team’s research doesn’t pose an immediate security risk. It isn’t likely to be applied immediately to viruses that infect humans. Those viruses are much larger and more complex than the ones synthesized in the experiment.
“When people hear that a virus was generated by AI, a lot of people are going to be terrified,” said Peter Koo, a computational biologist at Cold Spring Harbor Laboratory. “But human viruses are so different.”
Samuel King, a graduate student at Stanford University and an author of the study, said the team sought a deeper understanding of the DNA sequences that form all life on Earth. “This is what has produced all of the beautiful biological diversity around us,” King said.
The scientists developed the AI models for their study in collaboration with the Arc Institute, a research organization in Palo Alto, Calif. Just as consumer chatbots including ChatGPT are based on large language models that analyze mathematical relationships between words in text to create new text strings, their models, called Evo 1 and Evo 2, trained on genetic sequences to predict the next DNA base. For safety, they only trained the models on DNA from viruses that can’t infect humans.
King said they wanted to see if the AI could understand DNA in a way sophisticated enough to be able to capture all of the complex components of a genome that need to be orchestrated together. It turns out, they found, AI can.
The researchers used their model to generate genomes based on the template virus and then synthesized the DNA and put the AI-generated viruses into bacterial cells. If the bacterial cells died, their viruses were working. They tested about 300 of these designs.
Around 3 a.m. one night, they observed one of their viruses successfully replicating, and infecting and killing bacterial cells. “We were just very happy,” King said. Ultimately 16 of their viruses were functional.
Independent experts in computational biology and biosecurity said the research marks a major achievement in the field. It is the first time someone has used generative AI to redesign a virus, said Kevin Esvelt, an associate professor at the Massachusetts Institute of Technology.
It is possible that one day, he said, someone will use a similar approach to make viral proteins or viruses themselves that infect people and evade vaccine protection or natural immune responses.
“That tool will allow people to create new variants like Covid that will spread to infect most people,” said Esvelt. “And that will be bad. We should not do that.”
For now, Hie said, it would be considerably easier for a would-be bioterrorist to produce one of the many pathogens that already are 100% guaranteed to be deadly. “If I were a bad actor trying to design a pathogen or something to do harm, I would not use AI,” he said.
Write to Georgia Wells at georgia.wells@wsj.com and Alex Janin at alex.janin@wsj.com