
Stanford researchers fed a generative AI model millions of genetic sequences and asked it to generate new viruses starting from a short DNA snippet from a naturally occurring virus. Sixteen of the roughly 300 candidates it produced actually worked, infecting and killing bacteria in laboratory tests — a result that has triggered a fresh round of alarm among biosecurity experts who say the safeguards meant to catch dangerous DNA orders are full of holes.
How the AI Built Working Viruses From Scratch
The Stanford team, led by first author Samuel King along with researcher Brian Hie, trained generative AI models on millions of DNA sequences pulled from naturally occurring viruses, according to the WBUR program On Point with Meghna Chakrabarti. The model started with a short DNA snippet from bacteriophage ΦX174 and generated the rest of around 300 possible new bacteriophage genomes — viruses that infect bacteria rather than humans — and the team tested all 300, ultimately confirming 16 as viable, according to the same report. The researchers published their findings in the journal Science, with King serving as lead author on the paper, per Stanford Report.
King has said the team deliberately used E. coli C, a safe laboratory strain, as the host for the engineered phages targeting ΦX174 — notable as the first DNA genome ever sequenced, per the same report.
Bacteria, Not People — For Now
The viruses the Stanford team built attack bacteria, not humans, and Tom Inglesby, director of the Center for Health Security at Johns Hopkins Bloomberg School of Public Health, told WBUR the team's target was a harmless laboratory strain of E. coli, and that researchers voluntarily avoided training their tool on human-infecting viruses.
Esvelt said AI-designed phages could potentially help overcome antibiotic resistance by infecting bacterial strains that have grown resistant to existing treatments. Separately, Science reported that a mixture of the designed phages rapidly overcame E. coli strains that had become resistant to the natural phage ΦX174, while a comparable mixture of naturally sourced phages could not — a result Stanford Report said points toward engineered phage cocktails as a possible new class of antibiotics.
A 1918 Flu Experiment That Exposed a Loophole
To illustrate just how porous current safeguards are, Esvelt described an experiment in which his lab ordered pieces of the 1918 pandemic influenza virus from 38 DNA synthesis companies using a pseudonym and an office address, according to WBUR. Companies shipped the genetic material back from 36 of the 38 orders, Esvelt said.
Inglesby explained that federal policy currently requires only recipients of federal funding to buy from DNA synthesis companies that screen orders and customers. The policy requires screening sequences of 50 base pairs or longer, and Esvelt made a 30-base screening algorithm.
Industry Voices and Dissenting Views
Not every expert treats AI-designed genomes as the top threat. Tom Ellis told The Guardian that designing a genome for a full virus or bacterium with AI is a less likely pathogenic threat than simply taking existing pathogens and making gain-of-function changes to them, calling the latter path much easier and much more likely to produce a real danger. Hie, the Stanford researcher, made a similar point to Stanford Report, arguing that existing pathogens present a greater risk than potential AI designs specifically because they are easier to access and produce.
Inglesby and Hanke warned of biosecurity risks. Some AI companies have introduced voluntary safety protections.
Esvelt Points to a Darker History — and Possible Defenses
Esvelt also drew on historical precedent, noting that the Soviet Union weaponized smallpox and may have engineered vaccine resistance into it, and that the United States believes Russia currently has an active biological weapons program, according to his comments to WBUR. He said pandemic-proof protective equipment has not been stockpiled sufficiently for critical workers, leaving a gap he considers urgent regardless of where the next threat originates.
On the defensive side, Esvelt pointed to far-UV light, a technology with a wavelength of about 220 nanometers that can kill airborne pathogens without harming people. He said far-UV light is approved at levels that can eliminate well over 90 percent of airborne bacteria within a minute and could reduce infection-causing bacteria and viruses passing between people standing about a meter apart. Deploying such protective technologies, he said, could help prevent respiratory-transmitted illnesses including colds, flu, and COVID — diseases that Esvelt said cause hundreds of billions of dollars in lost productivity every year.
Legislative Push Faces a Narrow Window
Inglesby told WBUR that Congress is moving toward legislation that would require DNA synthesis screening even for people without federal funding, and that the proposal could conceivably pass during the current lame-duck session. Inglesby estimated that screening costs about a penny on the dollar for the company doing the work — a modest price, he argued, for intervention against high-end AI risks that need not slow beneficial technology.
He was careful to frame the stakes on both sides of the ledger. Inglesby said AI could help discover proteins with useful properties, including more effective antibiotics, uncover experimental pathways that may aid cancer research, and make medicine development faster, cheaper, and more effective overall. Esvelt, for his part, said he remains skeptical that current AI could cause human extinction in the near term, even as he singled out a mirror-image bacterium as one potential extinction-level mechanism worth taking seriously.









