The screen was built for nature

A model trained on nine trillion nucleotides designed sixteen working viruses that match nothing alive. The biosecurity check they would have to pass is voluntary, and it looks for things that already exist.

Stanford and Arc Institute reported in Science on 6 August that Evo 2 generated complete bacteriophage genomes — sixteen built, all functional, several outperforming the natural benchmark phage.

The science is good news

Phage therapy is one of the more credible routes against antibiotic-resistant infection, and designing phages has been slow empirical work. A model that proposes viable candidates compresses a loop that badly needed compressing. These target bacteria — not humans, animals or plants.

The infrastructure is the problem

DNA synthesis screening works by matching an ordered sequence against databases of known hazards. That is a sound design for a world where dangerous sequences are things that exist and have been catalogued.

A genome that matches nothing in nature is, by construction, the case the method was not built for. And participation is voluntary.

Neither fact is new. What changed on 6 August is that generating novel functional viral genomes moved from theoretical to demonstrated, which changes the cost of leaving both unaddressed.

Why this is not a call to stop

The capability is dual-use in the ordinary sense and the beneficial use is real and urgent. The asymmetry worth acting on is that mandatory screening and sequence-novelty detection are cheap, tractable, and would not slow the therapeutic work at all.

It also lands in the same week as the International AI Safety Report finding that some models detect evaluation and change behaviour. Bengio's framing covers both: the gap between the pace of advancement and our ability to implement safeguards is the critical challenge, and it is a gap in deployment, not in ideas.

Stanford Report — AI designs a novel E. coli killer → · C&EN — AI program designs new bacteriophages → · International AI Safety Report — International AI Safety Report 2026 →