Could it be true?
Could AI Actually Design a Dangerous Virus?
A year ago, that question belonged comfortably in science fiction. It doesn’t anymore.
There are sentences you expect to write as a thriller author and sentences you would rather not write as a former doctor.
Here is one of the latter:
Artificial intelligence has designed functioning viruses.
Not computer simulations of viruses. Not pretty molecular pictures on a screen. Not theoretical genetic sequences that might work if somebody were foolish enough to try them.
Actual viruses.
In 2026, researchers at Stanford University reported using AI to generate complete genomes for bacteriophages, viruses that attack bacteria. They synthesized nearly 300 of the designs. Sixteen produced viable phages.
They worked.
Before anyone starts digging a bunker in the garden, there is an important qualification. These viruses were designed to infect E. coli, not humans. They weren’t airborne pandemic pathogens. They weren’t biological weapons. They weren’t created to wipe out humanity before lunch.
In fact, bacteriophages could be enormously useful. As antibiotic resistance grows, viruses that kill dangerous bacteria may eventually become valuable medical tools.
So this research is potentially very good news.
Unfortunately, it also demonstrates something we should probably pay attention to.
AI can now help write biology.
And once you can write biology, the obvious question is what else you might eventually be able to write.
From reading DNA to writing it
For years, artificial intelligence in biology was mainly discussed as an analytical tool.
Give an AI a mountain of biological data and ask it to find patterns humans might miss. Predict how a protein folds. Search for possible drugs. Compare genetic sequences. Identify promising molecules.
That is powerful, but the AI is essentially reading biology.
Generative AI changes the relationship.
Instead of asking, “What does this sequence do?”, we can increasingly ask, “Can you design a sequence that does what I want?”
That is a profound difference.
The Stanford work used genome language models, including Evo 2, to generate complete bacteriophage genomes. The researchers started with an existing, well-understood phage called ΦX174, so this wasn’t an AI sitting alone in a darkened room inventing life from nothing.
There were scientists, laboratory equipment, synthesis, experiments and many failures.
That last point matters.
Of the nearly 300 generated designs that were synthesized, only sixteen produced viable viruses. Biology remains wonderfully stubborn. A DNA sequence that looks convincing to a computer doesn’t necessarily become a functioning organism.
Current AI also remains surprisingly poor at predicting some of the effects of genetic mutations. Recent research on viral genomes has shown just how difficult it can be to move from genetic sequence to reliable biological behaviour.
So we are nowhere near typing:
“Please design an airborne virus capable of infecting humans.”
And receiving a neat little file called Pandemic_Final_v3 five seconds later.
But that isn’t the part that worries me.
The worrying part is what happens when failure becomes part of the system.
What happens when the AI can try again?
Nature doesn’t design organisms perfectly on the first attempt.
It experiments.
Mutation produces variation. Selection keeps what works. The process repeats.
Now imagine applying something similar to artificial intelligence.
The AI produces hundreds of biological designs.
They are synthesized.
An automated laboratory tests them.
Most fail.
A few work.
The experimental results return to the AI.
It learns which changes produced which effects and generates another set.
Those are tested.
The results return again.
Get Sophy Smythe’s stories in your inbox
Join Medium for free to get updates from this writer.
Design. Build. Test. Learn.
Repeat.
This isn’t fantasy. Researchers already use versions of this cycle in synthetic biology, and laboratory automation is making the process faster. AI, robotics and biological engineering are gradually moving closer together.
That changes the important question.
It may not matter whether an AI can design the perfect virus on its first attempt.
What matters is whether it can learn how to design a better one on its hundredth.
At present, there are major barriers.
Human viruses are vastly more complicated than simple bacteriophages. Transmission, immune escape, host range, replication, disease severity and environmental stability interact in ways we still don’t fully understand.
Changing one characteristic may destroy another. A mutation that improves transmission might reduce virulence. Something that works beautifully in a computer model may fail completely inside a living organism.
Biology has an irritating habit of refusing to behave like software.
For the moment, that is reassuring.
But “AI can’t reliably do this today” and “AI will never be able to do this” are very different statements.
The people responsible for biosecurity certainly aren’t treating the second statement as safe.
NIST has warned that generative biological models could potentially be misused to modify dangerous biological properties. RAND has studied how AI might assist attempts to alter viral capabilities. And in September 2026, Anthropic reported cases in which people had already tried to use its AI systems while researching potentially dangerous biological work, including experiments involving avian influenza.
That doesn’t mean an AI created a bioweapon.
It didn’t.
But somebody has already thought of asking.
And that may be the least surprising part of this story.
The question we should be asking now
Technology doesn’t have intentions.
People do.
The same AI capable of designing a virus that destroys antibiotic-resistant bacteria might eventually help us develop new vaccines, cancer treatments or therapies for diseases we cannot currently cure.
That would be extraordinary.
It might save millions of lives.
But the technology that makes biological design faster, cheaper and more accessible doesn’t automatically ask whether the person using it is trying to cure a disease or create one.
That distinction has to come from us.
And history doesn’t offer much comfort when powerful technology becomes widely available.
The immediate danger isn’t that some superintelligence suddenly decides to exterminate humanity. The more realistic concern is much less cinematic: biological design tools become steadily more capable, laboratory automation becomes easier to access, and knowledge that once required years of specialist training becomes available through an AI assistant.
Each individual improvement may look harmless.
A better genome model.
Cheaper DNA synthesis.
More capable laboratory robots.
An AI that understands experimental results.
Software that proposes the next experiment automatically.
None of these creates a pandemic.
But connect enough of them and something changes.
The barrier between having a dangerous biological idea and being able to test it begins to shrink.
That is why I think the next twelve months will be fascinating — and uncomfortable.
I don’t expect an AI-designed human pandemic virus to appear next year. There is no scientific evidence that allows us to make that prediction.
I do expect AI-generated biological systems to become more sophisticated. I expect automated laboratories to become more capable. I expect the design-build-test-learn cycle to become faster. And I expect governments, AI companies and scientists to spend considerably more time discussing who should be allowed to use these tools and what they should be allowed to ask them to do.
Because the most important scientific question has quietly changed.
A few years ago, we could ask:
Can AI design a virus?
In 2026, the answer is no longer simply no.
So now we have to ask a much more uncomfortable question:
What kind of virus could it design next?
When I wrote The Hellraiser, I invented Protocol Eden: an airborne, AI-designed virus that can change, learn and ultimately make decisions about who should live and who should die. I deliberately pushed the science beyond what was possible because that is what thriller writers are allowed to do. Protocol Eden is still fiction, and I very much hope it stays there. But while writing the novel, I thought I was moving today’s science several years into the future. Looking at what has happened in 2026, I am beginning to wonder whether I moved it far enough and I’m asking myself:
Could it be true?
#ArtificialIntelligence #Biosecurity #SyntheticBiology #AISafety #Biotechnology
