AI can now help make “deepfake” biological viruses and won’t trigger our immune systems.
Universal Images Group via Getty Images
Recently, I had a conversation with Eric Nguyen, co-founder and CEO of Radical Numerics. He is one of the innovators behind teaching artificial intelligence to write DNA, aiming to combat cancers and other diseases through precision DNA editing and personalized medicines. However, during our discussion, he highlighted a threat that seemed more akin to a Hollywood cyber-thriller than academic biology: the concept of a deepfake virus.
Unlike digital threats targeting devices or finances, a deepfake biological virus poses a health risk by infiltrating the body’s cells. This virus is disguised in such a way—deepfaked—that it eludes detection by the immune system, appearing unlike any known virus.
“It’s possible to design DNA to mimic a virus while rearranging its genetic code such that it evades existing detection systems,” Nguyen explained in a recent NEXT podcast. “This involves modifying the sequence to maintain its functionality without matching previously identified viruses.”
This re-engineered virus, crafted by AI, retains its harmful effects while concealing its genetic sequence from both artificial and natural screening systems.
I wondered if this was a problem for the future. Perhaps next year, or the next decade?
Well, the future is here.
On August 6, researchers from Stanford and the Arc Institute released a paper in Science detailing how they utilized generative AI to create functioning viruses from scratch. These are the first machine-generated genomes that have not been identified in nature. The team created hundreds of potential genomes, synthesized nearly 300 in the lab, and successfully brought 16 to life as bacteriophages that infected and destroyed E. coli.
A framework for AI-guided bacteriophage genome design
Science
The AI models used were Evo 1 and Evo 2, developed by Eric Nguyen and others at the Arc Institute. Thus, Nguyen witnessed his own creation confirm the concept of deepfake viruses shortly after discussing it with me.
What the Stanford team actually built
Take a moment to relax. The Stanford team did not create a superbug. Instead, they designed bacteriophages—viruses that target bacteria, not humans. With antibiotic-resistant infections causing approximately two million deaths annually, phage therapy provides a valuable tool for medical scientists.
Developing better phages on demand can be highly beneficial.
Moreover, the team conducted the experiment responsibly. They removed human, animal, plant, and fungal virus sequences from their training data, preventing the model from learning to construct viruses that could infect humans or plants. Their work was conducted in secure facilities.
Yet, the potential of this technology is evident.
The Science paper demonstrates that AI can now generate bacteriophages, creating complete and functional virus genomes from a text prompt and training data. This is no longer science fiction. Crucially, the method is adaptable. Target it at phages, and phages are produced. Direct it elsewhere, and different results will emerge. The absence of more dangerous outcomes is solely due to the researchers’ restraint.
This restraint cannot be assumed of all research teams globally.
Deepfaking viruses with AI
And therein lies the issue.
Some safeguards do exist. DNA synthesis companies employ screening software to identify and halt the shipment of dangerous sequences.
However, as Nguyen noted, AI can create genetic sequences that bypass detection systems. By altering the genetic code while maintaining its function, pathogens can disguise themselves, much like they evade our immune systems.
The result is the same lethal payload with a different signature. This is a biological deepfake.
Nguyen explained, “AI biological foundation models, which are trained to generate new protein and DNA sequences, can be intentionally used to design sequences that circumvent detection systems.” He added, “This is a growing concern, especially at the national security level.”
Such possibilities keep biosecurity experts awake at night. Nguyen points out that our capacity to design is advancing faster than our ability to defend. The three pillars of biodefense—early outbreak detection, attributing the source, and rapidly manufacturing countermeasures—are areas where, according to him, we are lagging significantly.
Rules, regulations, and laws don’t exactly cover this
One might assume there is a regulatory framework for this. There is, but only partially.
The U.S. recently banned federally funded gain-of-function research. However, this framework was designed for a scenario where an existing virus is modified by a scientist. It says little about AI-generated novel genomes.
Moreover, it’s merely a guideline for federal funding, not a comprehensive law. It applies to a single country, not the entire world.
As biosecurity researchers Thomas Inglesby and Moritz Hanke noted following the release of the Science paper, the capability to create viral genomes with generative AI now exists. However, the governance needed to safely manage it does not.
Of course, this technology is powerful for good too
In reality, this technology holds immense potential for positive impact.
AI’s ability to read and write DNA could enable earlier disease detection, expedite the development of personalized medications, and facilitate the rapid production of antivirals during pandemics. These are significant benefits.
However, there are also considerable risks.
Interestingly, Nguyen has reconsidered his views on openness as the stakes have increased. While Evo was released openly due to the public availability of DNA training data, newer, more advanced models from Radical Numerics are being withheld for safety reasons.
The stark reality is that our global health detection systems were designed for an era when creating a virus was challenging. Now, it’s become much easier, and that poses a danger.
The 16 phages in the Science paper are harmless to humans. Yet, the machine that enabled their creation can achieve other feats more swiftly than we can collectively respond.
Pandora’s box is open.
As this remains true, the only viable path is to continue developing this technology to counter those who might exploit it maliciously, hoping to establish effective and rapid defenses against future AI-designed deepfake viruses. Nguyen suggests we can build an AI system capable of quickly designing medicines to combat new threats as they arise.
It seems we will indeed need it.
