There you have it! AI just designed 16 viruses that don't exist in nature, and they work.
August 30, 2026 · 0 likes · 0 comments
Cybersecurity AI
There you have it! AI just designed 16 viruses that don't exist in nature, and they work.
Stanford and the Arc Institute fed a generative model called Evo trillions of genetic letters, asked it to invent new bacteriophages, and synthesized the designs into real DNA.
16 of them infected and killed E. coli in the lab.
Some of them killed BETTER than the natural virus they were based on. Including strains that had already evolved resistance.
Drug-resistant infections killed 4.7 million people in 2021. In the U.S. alone, 35,000 dead a year. Antibiotics are running out of road. On-demand, AI-designed phages could be a real weapon against the bugs that no longer respond to anything.
That's the good news. Now the part in the same journal issue that nobody wants on the slide.
The Johns Hopkins biosecurity team wrote it plainly: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not."
The capability shipped. The rulebook didn't.
The model that did this — Evo 2 — is fully open-source. Code, weights, training data, all public. The Stanford team was responsible and stripped human-infecting viruses out of its training set. But open-source means the NEXT team doesn't have to.
No federal agency has announced a single new rule in response. The recommendation to keep dangerous viral sequences out of future models is a proposal, not a policy.
I run AI systems in production every day. Here's the thing about a tool this powerful: it does exactly what its guardrails allow — nothing more, nothing less. The Stanford team built good guardrails. The problem is we're relying on every future team to volunteer the same ones, with no floor underneath them.
You don't leave that to goodwill. You don't leave it to whoever downloads the weights next.
We are getting very, very good at building capability. We are barely trying on containment.
We broke this down on unbiasedheadlines.com, cross-spectrum sources, no outrage algorithm, actual facts you can check: https://lnkd.in/ehW3MAhM
Cure or catastrophe — same tool. Who's writing the floor before the next model gets built?
Stanford and the Arc Institute fed a generative model called Evo trillions of genetic letters, asked it to invent new bacteriophages, and synthesized the designs into real DNA.
16 of them infected and killed E. coli in the lab.
Some of them killed BETTER than the natural virus they were based on. Including strains that had already evolved resistance.
Drug-resistant infections killed 4.7 million people in 2021. In the U.S. alone, 35,000 dead a year. Antibiotics are running out of road. On-demand, AI-designed phages could be a real weapon against the bugs that no longer respond to anything.
That's the good news. Now the part in the same journal issue that nobody wants on the slide.
The Johns Hopkins biosecurity team wrote it plainly: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not."
The capability shipped. The rulebook didn't.
The model that did this — Evo 2 — is fully open-source. Code, weights, training data, all public. The Stanford team was responsible and stripped human-infecting viruses out of its training set. But open-source means the NEXT team doesn't have to.
No federal agency has announced a single new rule in response. The recommendation to keep dangerous viral sequences out of future models is a proposal, not a policy.
I run AI systems in production every day. Here's the thing about a tool this powerful: it does exactly what its guardrails allow — nothing more, nothing less. The Stanford team built good guardrails. The problem is we're relying on every future team to volunteer the same ones, with no floor underneath them.
You don't leave that to goodwill. You don't leave it to whoever downloads the weights next.
We are getting very, very good at building capability. We are barely trying on containment.
We broke this down on unbiasedheadlines.com, cross-spectrum sources, no outrage algorithm, actual facts you can check: https://lnkd.in/ehW3MAhM
Cure or catastrophe — same tool. Who's writing the floor before the next model gets built?