Scientists Create First Viruses Designed by AI, Raising Global Biosecurity Alarm

The ability to write the genetic code of a living virus using artificial intelligence now exists—but the global safeguards to prevent that technology from being misused do not. In a milestone that bridges generative artificial intelligence and synthetic biology, researchers at Stanford University and the Arc Institute have successfully designed and synthesized 16 fully functional viruses from scratch. The breakthrough, detailed in the journal Science, marks the first time generative models have been used to design genome-scale sequences that function as viable viruses in the real world, extending AI-assisted biology far beyond the design of individual proteins.[2][3]

The viruses in question are bacteriophages, a specific class of virus that exclusively infects and hijacks bacteria. To create them, the research team, led by Dr. Brian Hie and Samuel King, utilized genome language models known as Evo1 and Evo2. Much like how large language models learn the patterns of human speech to generate coherent text, these biological models were trained on the genetic sequences of two million naturally occurring bacteriophages. By learning the evolutionary grammar of viral DNA, the artificial intelligence was able to recombine those patterns and propose entirely new genetic sequences that do not exist in nature.[1][2]

Generating the code was only the first step; the designs then had to be physically manufactured and tested. The scientists synthesized nearly 300 of the AI-generated genomes in the laboratory and inserted the synthetic DNA into bacterial cells. The bacteria read the artificial genetic code and churned out the novel bacteriophages. While the overall yield was relatively low, 16 of the bacteriophages proved to be viable. When mixed into a cocktail, these AI-designed viruses successfully infected and destroyed strains of E. coli that had already developed resistance to naturally occurring phages.[1][2]

The clinical implications of this success are profound. Antimicrobial resistance is a rapidly escalating global health crisis, rendering many conventional antibiotics ineffective against persistent superbugs. Phage therapy—the practice of using targeted viruses to kill resistant bacteria—is widely considered one of the most promising alternatives. The Stanford team’s research demonstrates that artificial intelligence could eventually be used to rapidly design and tune custom phages tailored to defeat specific, drug-resistant bacterial strains, potentially saving millions of lives and transforming how modern medicine approaches untreatable infections.[1][3]

However, the achievement has triggered immediate alarm among biosecurity experts who warn that the technology is advancing faster than the regulations needed to govern it. In a companion perspective piece, researchers Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security cautioned that the governance required to safely steer generative viral design is currently absent. They noted that while the Stanford team deliberately excluded the genetic code of human, animal, and plant pathogens from their AI's training data, the underlying capability to compose viral genomes from scratch could theoretically be misused to engineer dangerous new pathogens.[1][2]

The primary vulnerability identified by security professionals lies not in the software models themselves, but at the point of physical manufacturing. Currently, researchers and individuals can order custom-made DNA sequences from commercial synthesis providers online. While many of these companies voluntarily screen orders against databases of known dangerous pathogens, these checks are not universally mandated by law in most jurisdictions. Because an entirely novel, AI-generated sequence might not match any known threat in existing databases, it could potentially evade current screening protocols, allowing bad actors to obtain the physical building blocks of a synthetic virus.[1][3]

Despite these urgent warnings, some experts argue that the immediate threat of AI-generated human viruses is being overstated. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, acknowledged the impressiveness of the Stanford research but noted that bacteriophages possess some of the smallest and simplest genomes in biology. He emphasized that designing a complex virus capable of infecting humans entirely from scratch remains extraordinarily difficult. According to Ellis, the far greater and more immediate biosecurity threat comes from bad actors using traditional methods to edit and enhance existing, well-understood pathogens, rather than relying on generative AI to invent new ones.[1][3]

Nevertheless, the consensus among both synthetic biologists and security professionals is that the regulatory landscape must adapt rapidly to the realities of generative genomics. Proposals to mitigate the risks include mandating universal, legally binding screening for all commercial DNA synthesis orders and requiring developers of biological AI models to consult with biosecurity experts throughout the design process. As the boundary between digital code and biological life continues to blur, the defining challenge for the scientific community will be harnessing the immense therapeutic potential of artificial intelligence without inadvertently democratizing the creation of biological threats.[1][2][3]