Scientists just used AI to design a working virus from scratch.

Stanford worked with an AI model to construct a virus genome from beginning to end, built one in the lab and used it on bacteria that were resistant to antibiotics to kill them. This has never been done like this, end to end, before and the output it produces is truly practical.

It has also posed a question that nobody actually really knows the answer to yet: who is watching what these tools will be used against next.

The model created the entire genome, not simply one gene

Brian Hie and his coworkers trained two AI systems Evo 1 and Evo 2 to predict DNA sequences in the same way that a language model predicts subsequent words in sentences.

It started with a natural virus called ΦX174, which is a bacteriophage that infects E. coli and carries just under 6,000 base pairs of genetic code, mere flecks compared to the roughly 3 billion in a human genome, and generated from scratch 302 candidate genomes via the model. All of them were built and tested by researchers.

Sixteen of these turned out to be active, working viruses and a few were even better than the natural virus they came from at killing E. coli. The cocktail, consisting of AI-designed phages, conquered resistance in two bacterial strains that resisted natural phages entirely.

But no one trained it on anything destructive to a human.

The design of safety here was intentional. Using only bacteriophages (which, by definition, can infect and replicate only in bacteria), Evo's training data excluded any kind of virus capable of infecting humans, animals or plants.

It is an honest-to-goodness guardrail, but also a very thin one. "This is literally the smallest and easiest genome to make," said Tom Ellis, a synthetic genome researcher at Imperial College London. His actual concern runs the other way: taking a pre-existing and dangerous virus and making relatively minor changes to it would still be easier and much more feasible as a biological threat than designing a new one from scratch.

The debate that no one has ever actually resolved

Two Johns Hopkins biosecurity researchers, writing alongside the study, put it plainly: with generative AI now able to write viral genomes, we lack the governance to safely contain that.

Their concern isn't this specific outcome, it's the capability in general, since a more permissive version of the same model, trained on even more lethal genetic data, would be a very different beast. Filippa Lentzos, a biosecurity researcher at King's College London, contended that rather than limiting the AI model itself, the real fix is to tighten up the layer where genetic designs are physically produced as DNA, alongside existing research review and lab biosafety standards.

Naturally, no one is saying this study produced a threat. What it proved is that the base ability, an AI creating a functional viral genome from scratch, is no longer theoretical. When it comes to viruses, bacteriophages are some of the least harmful.

An open question is merely how difficult that would be with something else that isn't.

One of science fiction’s greatest writers warned us about AI. Could his ideas help us avoid a dystopian future?

What might a modern day equivalent of Isaac Asimov’s laws of robotics look like? Guided by the author, I propose the three laws of AI

AI Tool of the Week: Agent37

What does it do?

• One-click deployment, no server management. Agent37 lets traders launch a Hermes or OpenClaw trading agent without touching a VPS or handling any infrastructure setup themselves.

• Runs 24/7 without you babysitting it. Once deployed, the agent stays live around the clock, letting strategies execute continuously instead of only when you're at your desk.

• Built to keep costs down. It supports cost-effective models, so running an agent doesn't rack up excessive token or infrastructure costs the way a lot of always-on setups do.

🎁 Cryptopolitan readers get the first month free, valid for the first 100 users. Use code: CRYPTOPOLITAN

• Microsoft reveals nearly 70% of its AI revenue comes from a single customer: OpenAI.

• Elon Musk is building the world’s first 100 million square foot building.

• Tether purchased +14 tonnes of gold in Q2 2026, bringing total holdings to a record 146 tonnes, now worth ~$18.8 billion.

• The largest mortgage lender in the United States, United Wholesale Mortgage, fell to an all-time low after suffering its largest drop in history

• Thailand officially confirms 0% capital gains tax on Bitcoin and crypto for trades through licensed exchanges.

Join the Conversation!

We'd love to hear your thoughts and comments. Join our community and stay updated with the latest trends and discussions in crypto.

Twitter | Instagram | Telegram Channel | Linkedin | Facebook