The team built their AI model around a real-world template: ΦX174, a naturally occurring bacteriophage — a virus that infects bacteria rather than humans. Hie and King fine-tuned Evo 2 on roughly 14,000 viral genomes from the Microviridae family, the same family as ΦX174, a phage notable for being the first DNA-based genome ever fully sequenced, back in 1977. Given that starting point, Evo 2 suggested thousands of new DNA sequences, and researchers synthesized and tested nearly 300 of them for effectiveness against E. coli. Of those, 302 designs were synthesized, and 16 of the 285 tested proved fully viable — meaning they could actually replicate and function as living viruses despite not existing anywhere in nature.
The results exceeded what researchers had expected going in. Lab tests of Evo 2's designs for an E. coli killer exceeded expectations, and several of the AI-generated phages actually outperformed their natural counterpart. Some of the AI-designed versions proved more effective than natural phages at killing strains of E. coli that had developed antibiotic resistance, and researchers found that a cocktail of the generated phages rapidly overcame resistance in three different ΦX174-resistant E. coli strains, whereas ΦX174 alone could not overcome that resistance. Cryo-electron microscopy even confirmed that one generated phage relied on a DNA-packaging protein evolutionarily distant from anything seen in nature, underscoring just how far the AI strayed from its biological training data while still producing something that worked.
The practical motivation behind the research is the escalating global crisis of antibiotic resistance. As bacteria increasingly evolve to survive existing antibiotics, scientists have looked to bacteriophage therapy — using viruses to hunt and kill drug-resistant bacteria — as a potential alternative. The problem with a single phage, however, is that bacteria can eventually evolve resistance to it too. Hie explained that if bacteria gain resistance to a single phage, "it's game over for the medication," which is why a diverse, AI-generated library of phage variants could offer a more resilient therapeutic toolkit than nature alone provides.
Alongside the excitement, the study has triggered serious biosecurity concern among outside experts, precisely because it proves a capability rather than just a product. Fatemeh Vafaee, a biotechnology professor at the University of New South Wales, framed the real significance of the work as being less about the specific viruses created and more about the demonstrated capability itself, calling for stronger regulatory oversight of the technology. That concern was echoed directly by biosecurity specialists at Johns Hopkins University, who wrote in a commentary accompanying the Science publication that the findings raise "urgent biosafety and biosecurity questions." They argued the real issue is no longer whether generative viral genome design will exist, but whether the technology can be developed and used without enabling serious harm — and were explicit that efforts to design new viruses capable of causing disease "should not be pursued." One especially pointed warning noted a regulatory gap: no U.S. law yet forces DNA-synthesis companies to screen orders for AI-generated sequences like the ones Evo 2 produced.
Experts are careful to note that bacteriophages targeting bacteria are a far cry from viruses capable of infecting and harming humans, and that designing a dangerous human pathogen remains a substantially harder technical challenge. Still, the timing has amplified the unease: the phage breakthrough arrived amid separate reports of AI systems from major developers acting outside their intended instructions during security testing, and against a backdrop of renewed U.S. political scrutiny over dangerous virus research generally — including congressional hearings on the origins of Covid-19 and disclosures about U.S.-funded biological research abroad. Researchers have responded by stressing that they've built in safeguards and have made Evo 2 openly available specifically so the scientific community can develop matching defensive tools, arguing that transparency and safety screening now, while the capability is still nascent, is preferable to encountering it unprepared later.
Sources & further reading:
- Science — Generative design of bacteriophages with genome language models
- Stanford Report — AI designs a novel E. coli killer
- PubMed — Study abstract and citation record
- bioRxiv — Original preprint
- RT News — Original source article
Cloudflare Connect 2026
Join thousands of developers, architects, and security leaders from October 19-21 at Moscone West in San Francisco for our biggest Connect yet. A place for you to learn, network, and meet the team powering over 20% of the web.
Save $100 with code: CFAGENTS26
Expires Sunday, August 9 at 11:59 PM PST
