AI now spits out entire viral genomes, and Stanford says it’s time to talk risk

The AI world has been obsessed with protein design for years. Makes sense — proteins do the heavy lifting in biology. But a quieter, stranger project has been running in parallel: training large language models on raw DNA sequences instead of amino acid chains. These ‘large genome models’ learned to predict the next nucleotide the way ChatGPT predicts the next word. And the results are getting weird.

A team at Stanford fed one of these models a prompt — a snippet from a bacterial gene cluster — and it output something far more ambitious than a single protein. It spit out full genomes for bacteriophages, the viruses that infect bacteria. These aren’t entirely novel pathogens plucked from thin air. The generated viruses are closely related to existing ones. But they carry distinct genetic features that would be difficult to evolve through natural mutation alone. That’s the part that should make you pause.

Brian Hie, the Stanford researcher who led the work, isn’t sounding a full alarm — yet. But he’s pointedly arguing that the conversation about guardrails needs to start now. The model that built these phage genomes wasn’t designed to be a bioweapon toolkit. It was a tool for understanding microbial ecosystems. The implication is clear enough: if a model trained on bacteria can produce functional viral genomes, a model trained on vertebrate viruses — with some modest engineering — likely could too. This isn’t a hypothetical for the distant future. The capability is already partially baked.

Here’s what I find unsettling, and what the paper doesn’t quite spell out: the accessibility problem. You don’t need a wet lab to run these models. The computational barrier is dropping fast, and the DNA synthesis to bring a digital genome to life can be outsourced. The Stanford work is a proof of concept wrapped in a warning label. It’s a reminder that biological design tools have no off-switch built into their core architecture — only the intentions of the people who use them. For now, the viruses only infect bacteria, but the principles scale.