During a recent conversation with Eric Nguyen, co-founder and CEO of Radical Numerics, I learned about a groundbreaking yet concerning application of artificial intelligence (AI) in the field of biology. Nguyen, an expert in programming AI to construct DNA, aims to harness this technology for targeted cancer therapies and personalized medicine. However, he also highlighted a potential threat that sounds like it's straight out of a cyber-thriller: the emergence of "deepfake" biological viruses.
Unlike digital threats aimed at corrupting your devices or stealing personal information, these deepfake viruses are designed to infiltrate human cells undetected. They mimic normal viruses so effectively that our immune systems fail to recognize them. As Nguyen explained in a recent podcast, "One could design DNA to essentially function like a virus, but be able to obfuscate or intentionally basically switch the letters around … so that existing detection systems cannot actually notice that it’s a virus that they’ve seen before." This manipulation of genetic code allows harmful viruses to retain their dangerous capabilities while disguising themselves from our biological defenses.
Initially, I assumed this was a concern for the distant future—perhaps a decade away. However, it seems that timeline has rapidly shifted.
On August 6, a collaborative research team from Stanford and the Arc Institute published a groundbreaking study in Science, demonstrating that they could use generative AI to fabricate functioning viruses from scratch. These are the first fully artificial genomes never seen in nature. The team produced hundreds of potential genomes, created nearly 300 in the lab, and vivified 16 of them into bacteriophages capable of effectively infecting and destroying E. coli bacteria.
Interestingly, the AI frameworks employed in this research were Evo 1 and Evo 2, developed by Eric Nguyen and his colleagues. In a twist of irony, the very technology Nguyen warned could be weaponized had now produced a practical proof of concept just days after our discussion.
It's important to clarify some key points about the Stanford team's work. They did not create a dangerous superbug; rather, they engineered bacteriophages—viruses that specifically target bacteria, not humans. Such technology can be a powerful tool against antibiotic-resistant infections, which claim roughly two million lives annually. The researchers conducted their experiments responsibly, excluding sequences from viruses known to infect humans, animals, plants, or fungi, and worked in secure environments. However, the implications of their findings are significant: the ability for generative AI to create viable viral genomes indicates we have moved beyond the realm of science fiction. The inherent flexibility of these generative models means they can potentially produce an array of different viruses on command, and not all researchers may exercise the same caution.
This capability raises pressing concerns. Although there are some precautionary measures in place—DNA synthesis companies often implement screening software to flag hazardous sequences—the systems may not be foolproof. Nguyen pointed out that AI can develop genetic structures that cleverly bypass these detection mechanisms. By altering the sequence of genetic material while preserving its function, a pathogen could evade not only the filters established by the industry but also the defenses of our immune systems. This makes for a biological deepfake: a virus with the same lethal potential but an altered genetic signature.
Nguyen emphasized that as we advance our capacity for genetic design, our ability to safeguard against potential biological threats isn't keeping pace. He identified three critical components of biodefense: early outbreak detection, accurate attribution of the virus’s origin—whether natural, laboratory-based, or an intentional release—and the rapid development of countermeasures. He expressed concern that we are lagging significantly in all these areas.
As for regulations, you might expect there to be a robust framework overseeing this new frontier. However, while the U.S. has placed restrictions on gain-of-function research funded by federal sources, such regulations primarily address modifications of existing viruses, not the development of entirely new genomes via AI. Moreover, they apply to federal projects and do not extend globally. Biosecurity experts Thomas Inglesby and Moritz Hanke noted the alarming reality: the technology to create viral genomes with generative AI now exists, but effective governance to manage its use has not yet been established.
