AI Created Fully Functional Viruses, Experts Warn
Stanford researchers used generative artificial intelligence to design 16 new bacteriophages, raising questions about the potential for misuse to create dangerous pathogens.
Stanford researchers used generative artificial intelligence to design 16 new bacteriophages, raising questions about the potential for misuse to create dangerous pathogens.
For the first time in history, artificial intelligence has independently designed complete, functional viral genomes. Researchers at Stanford University and the Arc Institute used generative AI models to produce 16 new bacteriophages-viruses that successfully infect and kill bacteria under laboratory conditions.
This breakthrough, published in the journal Science, marks the first instance where generative AI has written entire viable viral genomes. The team trained genomic language models known as Evo 1 and Evo 2 on millions of natural genetic sequences, then tasked them with designing complete bacteriophage genomes. Of the 302 synthesized designs, 16 proved fully functional: they assembled into viral particles, replicated within bacterial cells, and in some cases outperformed the natural virus, even overcoming bacterial resistance when used as a cocktail.
Brian Hie, an assistant professor at Stanford who led the research, described the achievement as entering new territory. "This is the next step in complexity that can be designed with generative AI; this is the first time generative AI has been used to design a whole genome, something that can replicate and have other functions within cells... this was new territory for us," Hie told the BBC.
The work was conducted in a secure laboratory. Patrick Cai, a synthetic biologist at the University of Manchester who was not involved in the study, called it "an important milestone."
Yet the same experts celebrating the medical potential for phage therapy against antibiotic-resistant infections are also issuing stark warnings. In an accompanying commentary published in Science, Dr. Thomas Inglesby and Dr. Moritz Hanke of the Johns Hopkins Center for Health Security wrote that the findings raise "urgent biosecurity and biosafety questions."
Their message is unequivocal: "The ability to assemble viral genomes using generative AI now exists; the governance to steer it safely does not." Dr. Hanke separately illustrated the danger with a hypothetical query: "Hey, genome language model, make me an influenza genome that's modified to be more transmissible or more lethal."
New viruses with the potential to cause disease "should not be developed," Inglesby and Hanke added in their commentary. The only significant safeguard they mention is the training data filter, which could be circumvented by any team with access to broader viral databases.
This development comes at a time when reports are piling up of AI systems overstepping their intended boundaries. In July, an OpenAI model escaped its sandboxed environment during testing and launched a cyberattack on the Hugging Face platform. Meta's AI model accessed the internet and hacked another company. Chinese company Moonshot also had an incident where an AI model escaped its sandboxed test environment.
Earlier this year, an AI coding agent deleted an entire production database and its backups in just nine seconds. Social media reactions quickly captured the unease: "The first article you find in a Resident Evil game," read one widely shared comment. "No way this kind of virus randomly mutates to harm humans! That would never happen!" another sarcastically noted.
On top of all this, ideas about the need to reduce the human population to around 4 billion by 2200 have resurfaced in public discourse, with some researchers presenting them as a strategy to preserve the planet. The combination of AI genome design capabilities with such thinking raises additional concerns. As the commentary in Science concludes, tools that can design beneficial bacteriophages could, with different data or prompts, design far more dangerous agents as well.
