Artificial intelligence has already changed the way we write software, analyze data and search for information. Now researchers are demonstrating something considerably more consequential: AI can help design biological systems that have never existed in nature.
Researchers at Stanford University, working with collaborators including the Broad Institute and Arc Institute, used a generative AI model called Evo 2 to design complete genomes for bacteriophages — viruses that infect bacteria rather than people. Scientists synthesized hundreds of the computer-generated designs and found that 16 of the tested designs were functional. Some were capable of killing E. coli, including strains resistant to naturally occurring bacteriophages.
That is an impressive scientific achievement, but I think the most important part of this story is being overlooked.
The breakthrough isn't simply that scientists created a new type of virus.
It demonstrates that generative AI can move from analyzing biology to designing functional biological systems.
That is a much bigger change.
From Predicting Biology to Designing Biology
Traditional biological research is often based on observing what already exists.
Scientists study organisms, sequence their DNA, identify useful characteristics and then attempt to modify or reproduce those characteristics.
AI introduces another possibility.
Instead of limiting researchers to biological designs already found in nature, a sufficiently capable model can explore enormous numbers of possible genetic sequences and identify designs that may satisfy particular biological constraints.
This is conceptually similar to what generative AI has already done in other fields.
A large language model doesn't simply retrieve sentences written by someone else. It can generate a new sentence.
An image model doesn't simply retrieve a photograph. It can generate a new image.
A genomic model can potentially do something similar with biological sequences.
The important difference is that biology has consequences in the physical world.
The Stanford Research Is Not About Creating a Human Virus
There is an important distinction that headlines can easily obscure.
The viruses created in this research are bacteriophages. Their biological target is bacteria, including E. coli. They are not human viruses such as influenza, HIV or SARS-CoV-2.
That makes the research considerably less alarming than the phrase "AI-designed viruses" might initially suggest.
At the same time, the research demonstrates a capability that didn't previously exist at this scale.
The researchers were able to generate complete viral genomes with AI and then test whether those designs actually worked.
That is the part worth paying attention to.
Why Antibiotic Resistance Makes This Interesting
There is also a very practical reason scientists are interested in AI-designed bacteriophages.
Antibiotic resistance continues to make certain bacterial infections increasingly difficult to treat.
Bacteriophages provide a completely different approach. Instead of using an antibiotic that chemically interferes with bacteria, a phage can specifically infect and destroy susceptible bacterial cells.
The problem is that bacteria can evolve resistance to phages too.
This is where AI-generated diversity could eventually become useful.
Instead of relying on a relatively small collection of naturally occurring phages, researchers could potentially use computational models to explore a much larger biological design space.
The Stanford experiments are an early demonstration of that concept. Some generated phages were able to kill E. coli strains that resisted the natural reference phage. A combination of generated phages also showed an ability to overcome resistance in several tested bacterial strains.
That doesn't mean AI has solved antibiotic resistance.
It does suggest that AI could become another tool for discovering biological treatments.
The Most Important Word Is "Functional"
One detail in this research deserves particular attention.
Generating a DNA sequence on a computer isn't the same thing as creating a working biological system.
An AI model can produce millions of theoretically interesting sequences. Most of them may be useless.
The researchers therefore had to determine whether the designs actually functioned when synthesized and tested.
That is an important distinction when evaluating AI's role in science.
We've become accustomed to generative AI producing convincing text, images and computer code. Biological systems are different.
A generated design ultimately has to survive contact with reality.
In this case, some of the designs did.
That's what makes the result significant.
This Is Similar to the Evolution of AI in Cybersecurity
There is an interesting parallel here for people working in IT and cybersecurity.
Early AI security tools primarily analyzed information.
They detected patterns, classified events and helped humans identify suspicious activity.
Modern AI systems are increasingly capable of taking actions.
They can write scripts, interact with APIs, investigate systems and perform multiple steps without waiting for a human after every operation.
The same general progression is now appearing in scientific research.
AI is moving from:
Analyze → Predict → Generate → Test
That last step is particularly important.
When AI-generated designs can be automatically tested in the real world, the feedback can be fed back into the computational process.
That creates a much faster research cycle.
The computer proposes a design.
The laboratory tests it.
The results provide new information.
The model proposes additional designs.
That could dramatically accelerate certain areas of scientific research.
And That's Where the Security Question Begins
Every powerful technology eventually raises the same question:
What happens when the technology becomes easier to use than the expertise required to use it?
That question is particularly important with generative biology.
Today, designing and validating biological systems still requires specialized equipment, expertise and laboratory resources.
AI does not eliminate those requirements.
But it can potentially reduce the amount of specialized knowledge needed to explore biological possibilities.
That creates both an opportunity and a risk.
The opportunity is faster development of medicines, diagnostics and biological research.
The risk is that the same computational capabilities could eventually be used for purposes that researchers never intended.
Biosecurity experts have already raised concerns about the implications of generative models being used to design biological systems.
AI Safety Cannot Stop at the Model
This is where I think the discussion needs to move beyond the usual "Is AI dangerous?" debate.
The model itself is only one part of the system.
A safer architecture requires controls around the model.
For example, organizations developing biological AI systems need to think about:
What biological information is available to the model?
Who is allowed to use the system?
What types of designs can it generate?
How are potentially dangerous requests identified?
How are generated sequences evaluated?
Who approves physical experiments?
How are laboratory activities monitored?
What happens when an AI-generated design produces an unexpected result?
These are fundamentally governance and security questions.
And they have a lot in common with enterprise cybersecurity.
The Network Security Analogy
For network engineers, there is a useful analogy.
You wouldn't put a new automated application directly onto a production network and give it unrestricted administrator access.
You would normally put controls around it.
You would authenticate users.
You would restrict permissions.
You would segment systems.
You would monitor traffic.
You would log activity.
You would have a way to shut the system down.
The same philosophy makes sense for increasingly autonomous scientific AI.
Don't assume the model will always behave exactly as expected.
Design the surrounding system so that unexpected behavior is contained.
That's a principle that applies equally well to AI-powered network automation and AI-powered biological research.
Should We Be Worried?
Yes — but probably not for the reason suggested by some of the more sensational headlines.
The Stanford research doesn't mean that AI has suddenly learned how to create dangerous human pathogens.
It demonstrates something narrower but potentially very important: AI can generate biological designs that are sufficiently coherent to produce functional bacteriophages that researchers did not find in nature.
That is a scientific milestone.
Whether it becomes a major biosecurity problem depends on what happens next.
Governments, researchers, AI developers and biotechnology companies now have an opportunity to establish sensible safeguards while the technology is still developing.
Waiting until a dangerous incident occurs would obviously be the wrong strategy.
The Medical Potential Is Huge
It's also important not to lose sight of the positive side.
The same technology generating legitimate biosecurity concerns could eventually help researchers develop treatments for diseases that are difficult to address using conventional methods.
AI could potentially help scientists search enormous biological design spaces much faster than humans can.
That could lead to better therapies, new antibiotics, more targeted treatments and improved understanding of biological systems.
The Stanford work is therefore not simply a warning story.
It's a demonstration of what happens when generative AI moves beyond text and images and begins interacting with the physical sciences.
My Take
The most significant part of this research isn't that scientists created 16 functional bacteriophages.
It's that AI-generated biological designs crossed the boundary between computer simulation and physical reality.
That's a major conceptual shift.
We've spent the last few years discussing whether AI can write convincing text, generate realistic images or produce useful computer programs.
The next phase is much more interesting — and much more complicated.
AI is increasingly being used to design things that can actually exist in the physical world.
That means the conversation about AI safety needs to expand beyond misinformation, privacy and cybersecurity.
It also needs to include biology.
The technology itself isn't inherently good or bad. The outcome depends on how it is developed, who can access it, what safeguards surround it and how carefully its real-world effects are monitored.
For now, the Stanford research is an impressive demonstration of what generative AI can accomplish in biology — and a useful warning that the era of AI-designed physical systems is arriving much faster than many people expected.
Sources & further reading:
- Science — Generative design of bacteriophages with genome language models
- Stanford Report — AI designs a novel E. coli killer
- PubMed — Study abstract and citation record
- bioRxiv — Original preprint
- RT News — Original source article
Cloudflare Connect 2026
Join thousands of developers, architects, and security leaders from October 19-21 at Moscone West in San Francisco for our biggest Connect yet. A place for you to learn, network, and meet the team powering over 20% of the web.
Save $100 with code: CFAGENTS26
Expires Sunday, August 9 at 11:59 PM PST