A Stanford PhD student used AI to design viruses from scratch. It actually worked.

Key points

• Stanford University PhD student Samuel King used a generative AI model in 2025 to generate genetic blueprints for novel microscopic viruses.
• The AI-designed viruses already kill bacteria in lab tests, a finding that could eventually offer an alternative to antibiotics.
• King was named one of MIT Technology Review's Innovators Under 35 in 2026.
• A live conversation between King and MIT Technology Review reporter James O'Donnell is scheduled for 16 October 2026 at 18:30 BST.

Samuel King didn't sit down to change biology. He sat down with a PhD problem and a generative AI model, the same category of technology behind tools like ChatGPT, which predicts what comes next in a sequence. His sequence wasn't words. It was DNA.

In 2025, King fed the model genetic information and asked it to propose blueprints for viruses that don't exist in nature. The model obliged. More surprising: when researchers tested the designs in a lab, the AI-invented viruses killed bacteria. That's not a metaphor. They physically destroyed bacterial cells.

Why does this matter for ordinary people?

Bacteria that resist antibiotics kill more than a million people a year globally. Viruses that target bacteria, called bacteriophages, are a long-studied alternative, but finding or engineering the right one has always been slow, expensive, and largely guesswork. King's method replaces some of that guesswork with a model that can propose thousands of candidate designs quickly.

It's not a cure for anything yet. King himself is clear that this is early-stage research, not a clinic-ready treatment. But it's the first time a generative AI has produced functional virus blueprints that work in tests, and that's the line researchers have been trying to reach for years.

Think of it this way. If you wanted to design a new key, you could try every shape at random, or you could train a model on every key and lock ever made and ask it to propose a shape likely to fit. King taught the model what known viruses look like genetically, then asked it to imagine new ones.

What happens next?

Turning these designs into treatments will take years of safety trials. The bigger shift is the method itself.

The biosecurity dimension is real. We've been covering it closely: our September story on Anthropic blocking scientists from extracting bioweapon guidance from Claude shows that regulators and AI companies are already in reactive mode. A tool that generates working virus blueprints will concentrate those concerns considerably. For now, King's work sits inside a university lab with institutional oversight, which matters.

Generative AI is also moving fast into protein and molecular design. Our 30 September story on Google DeepMind's SynthID Bio covered how the lab is watermarking AI-designed molecules so their origin can be traced. That kind of provenance tracking looks more urgent now.

MIT Technology Review is hosting a live conversation with King on 16 October 2026. The honest read: it's early, it's real. The method is the story, not the viruses themselves.