Introduction

Stanford and Arc Institute scientists used an AI model to invent brand-new viruses that had never existed in nature. Not by tweaking a natural virus, but by writing its entire genetic code from scratch. These AI-designed viruses, called bacteriophages, only infect bacteria (in this case, E. coli), not humans. When mixed together, the AI-made viruses did something impressive: they beat drug-resistant bacteria that a natural virus alone couldn’t defeat.

How did they do it?

The researchers trained an AI model called Evo 2 on millions of real genomes from across the tree of life, teaching it the “rules” of what makes a genome biologically functional. Starting from just a small snippet of a known virus (ΦX174), Evo 2 generated thousands of complete, novel genome designs on a computer. Scientists then synthesized nearly 300 of those AI-written blueprints in the lab and found 16 that worked as living, functioning viruses, some even outperforming the natural original.

Pros vs. Cons

CON: Speed changes everything

Designing a functional viral genome by hand, through trial-and-error genetic engineering, has historically taken scientists years of painstaking work, because genes, regulatory switches, and structural elements all interact in ways that are hard to predict. With Evo 2, the AI generated thousands of candidate genome designs computationally in a fraction of that time, and researchers narrowed those down to viable viruses within a single study cycle. Now picture a state actor pursuing bioterrorism: with a well-funded lab, few legal restrictions, and a determined team, that same acceleration could plausibly compress a bioweapons timeline down to months rather than years. What’s actually stopping a hostile government? This study only demonstrates it for bacteriophages that can’t infect people, but the same generative approach, in principle, doesn’t know the difference between a virus that kills bacteria and one that infects humans.

The researchers themselves are careful to note that a digital design is not the same as a working bioweapon. You still have to synthesize it, get it to actually function, and test it, and most AI-generated sequences fail. But that caveat assumes normal constraints: modest funding, limited lab access, ordinary timelines. A state actor with unlimited funding, dedicated biosafety-level labs, and no ethical or legal restrictions, the position a country like the United States would be in if it decided to pursue this, would not face those same bottlenecks. The “it still takes lab work” reassurance is really a statement about resource constraints, not a hard technical ceiling.

PRO: Real medical upside

The clearest benefit is faster, more adaptable treatments for drug-resistant bacterial infections. Antibiotic resistance is already a major killer, and phage therapy which is using viruses that hunt and kill specific bacteria, has been explored for decades as an alternative. What’s new here is that AI can generate whole cocktails of genetically diverse phages on demand, which is exactly what you’d want against a bacterium that’s evolving resistance in real time: if the bug adapts to one phage, another in the mix can still take it down.

Why to be concerned

This isn’t a hypothetical worry. A separate, related study already shows the guardrails meant to catch misuse are shakier than assumed. Microsoft researchers ran a “red team” test in which they used AI protein-design tools to generate tens of thousands of variants of dangerous toxins like ricin, botulinum, and Shiga toxin, then checked whether DNA-synthesis companies’ safety screening software would catch orders for them. It didn’t, one screening tool reportedly missed the majority of the AI-redesigned toxin sequences, because the AI could “paraphrase” a toxin’s genetic code into something structurally similar but different enough to slip past the filters. Put the two stories together: a generative model that can write functioning genomes from scratch, and screening software that can be fooled by AI-generated variants of known toxins and the biosecurity gap looks less like a future risk and more like a current one.

My View

I think governments need to move now, either by restricting or tightly conditioning public funding for AI-biology research that touches genome design, or by mandating that any AI biodesign tool be paired with enforceable, updated screening requirements at every DNA synthesis vendor, not just the ones that volunteer to comply. At the same time, I don’t think we can pretend this is fully solvable: once a capability like this exists and the underlying models are openly published (as Evo 2 was), there’s no way to guarantee bad actors won’t find a way to misuse it, the goal isn’t to eliminate the risk, which isn’t realistic, but to raise the cost and difficulty of misuse as much as possible while the useful, life-saving applications are still allowed to develop.

References:

1. King et al., “Generative design of bacteriophages with genome language models,” Science (Aug. 6, 2026)
2. CNN, “AI creates 16 new viruses from scratch, showing promise for drug resistance and drawing warnings about potential for misuse” (Aug. 6, 2026)
3. Science, “Made to order bioweapon? AI-designed toxins slip through safety checks used by companies selling genes.”