Artificial intelligence can now generate complete viral genomes in the lab. That sentence alone has triggered headlines about “perfect biological weapons.” The reality is more specific, more technical, and more important to understand than the fear-driven framing suggests.
Researchers have shown that AI systems trained on genetic data can design new viruses and redesign known toxins in ways that slip past some existing safety checks. At the same time, those same tools are already being used to explore new medical treatments, including alternatives to antibiotics.
This article breaks down what scientists have actually demonstrated, where the risks truly lie, what safeguards are already in place, and why this is a biosecurity issue that deserves attention without exaggeration.
What Scientists Mean When They Say AI Can “Create Viruses”
When scientists say AI can “create viruses,” they mean that an AI model can design a complete genetic sequence that follows the rules seen in real viruses. The system does not build anything on its own. It generates DNA sequences that appear biologically plausible based on patterns learned from existing viral genomes.
This is a design step, not an automated creation process. The output is a digital genome that may or may not function when tested. Whether it becomes a working virus depends entirely on later laboratory validation, where most proposed sequences fail.
What is new is scale and novelty. AI can generate many full-length viral genomes quickly, including sequences that do not closely resemble any single known virus. That expands what researchers can test and makes it harder to predict behavior based on familiar reference genomes alone.
Why Bacteriophages Were Used as the Test Case
Bacteriophages are a practical proving ground for AI designed genomes because they let researchers test a full design to synthesis to validation loop with manageable safety and logistics. Many phage systems have decades of baseline data, standardized lab assays, and clear readouts for whether a design works, such as whether it forms plaques on a bacterial lawn and reproduces across multiple rounds. Phages also replicate quickly, which makes it feasible to evaluate many candidate sequences and learn what design choices tend to succeed or fail without tying up long animal studies or specialized clinical infrastructure.
They are also useful because the host is a defined bacterial strain that can be controlled tightly in the lab, which reduces experimental ambiguity when a design fails. That matters for AI generated sequences, where the key question is whether the proposed genome is internally coherent enough to function, not whether a complex host immune response changes the outcome. In short, phages let scientists stress test the core claim behind genome language models, namely that patterns learned from existing genomes can be used to generate new sequences that still behave like real viruses when built and tested.
The Real Concern: Dual-Use Research
Dual use research is work intended for legitimate scientific or medical goals that could also be repurposed for harm. With AI assisted biology, the concern is not that the technology directly creates dangerous agents, but that it changes who can explore biologically active designs and how quickly they can do so.
AI models trained on biological data can generate large numbers of plausible protein or genome designs with far less manual trial and error. That means early exploration, which once required deep domain expertise and time, can now be accelerated and partially automated. As a result, the gap between what a system can propose digitally and what oversight mechanisms are prepared to evaluate has widened.
An open access paper in PNAS Nexus examines this shift and describes it as a growing category of dual use capabilities of concern. The authors argue that the primary risk is not intentional misuse by most researchers, but increased uncertainty for monitoring systems that rely on recognizing known threats. When AI produces designs that are novel yet biologically plausible, traditional list based or similarity based controls struggle to assess intent or risk early in the pipeline. The paper emphasizes that governance approaches used for earlier dual use life science research can still apply, but only if they are adapted to focus on capability, scale, and access rather than specific sequences alone.
What the Science Study Actually Demonstrated
The Science study focused on a specific weakness in current biosecurity practice, namely how DNA synthesis providers screen genetic orders for potential misuse. The researchers tested whether modern protein design tools could take proteins already recognized as hazardous and redesign them so they would no longer be flagged by standard screening systems.
They found that many existing screening methods rely heavily on sequence similarity to known pathogens or toxins. When AI tools rewrote these proteins into new sequences that changed the underlying genetic code while preserving likely biological function, those redesigned sequences often passed through screening undetected. This showed that the issue was not a lack of oversight, but a mismatch between how screening systems define risk and how AI now generates biological novelty.
Crucially, the study did more than document the vulnerability. The researchers developed updated screening approaches that evaluate protein structure and functional features rather than sequence similarity alone. When applied, these methods substantially improved detection of AI redesigned sequences that would previously have gone unnoticed. The study demonstrates that AI exposes real but addressable gaps in biosecurity and that defensive systems can evolve alongside design tools when those gaps are identified and tested directly.
Why This Does Not Mean Bioweapons Are Suddenly Easy to Make
Designing a genome on a computer is an early research step, not a shortcut to creating a dangerous pathogen. Turning a digital sequence into a contagious human virus still requires specialized containment facilities, experienced personnel, repeated validation, and extended experimentation. The recent work involved small bacterial viruses that are far simpler than human pathogens, and even then, most proposed designs fail to become stable or predictable.
Concerns persist because some barriers are gradually lowering. Automation, cheaper DNA synthesis, and more capable AI models make exploratory work faster, even if they do not remove the need for expertise or infrastructure. This is why biosecurity focuses on choke points rather than assumptions about intent.
DNA synthesis screening remains one of the most effective of those choke points. Providers increasingly review both sequences and customers, and newer methods assess likely biological function instead of simple sequence matching. Industry standards and funding tied to vetted suppliers add another layer of constraint. Together, these factors mean AI is reshaping biological research, but it has not eliminated the practical and institutional barriers that prevent rapid or casual creation of bioweapons.
Global Efforts to Reduce Misuse Risk
Governments and international organizations are increasingly treating AI enabled biology as a distinct security issue rather than a subset of existing biotechnology policy. The focus is on understanding how advanced models change risk profiles and on building oversight systems that account for speed, scale, and accessibility, not just specific biological materials.
In the United Kingdom, the AI Safety Institute was created to test advanced models and evaluate potential misuse across sectors, including biology. At the same time, international initiatives are working to update biosecurity norms so that screening practices, risk assessment, and expectations for responsible research remain aligned as both AI and synthetic biology evolve.
Some proposals look beyond laboratories and suppliers, suggesting environmental monitoring such as wastewater or air sampling to detect signs of unauthorized biological activity. These approaches are still under discussion and raise practical and ethical considerations, but they reflect a broader shift toward layered detection and prevention rather than reliance on any single safeguard.
Why the Medical Upside Still Matters
The same capabilities that raise biosecurity concerns are also addressing long-standing limits in medical research. Designing and testing biological molecules is slow, expensive, and often constrained by trial and error. AI assisted models can narrow that search by proposing candidates that are more likely to work, allowing researchers to focus lab resources on fewer, better informed experiments.
In infectious disease research, this approach is already being used to explore new antibiotics, refine vaccine targets, and design bacteriophages that match specific bacterial strains more precisely. These are areas where traditional pipelines struggle, particularly as antibiotic resistance rises and many pharmaceutical companies retreat from antimicrobial development because of cost and low returns.
What matters most is that these tools can shorten early discovery timelines rather than replace clinical testing or regulatory review. For patients facing infections that no longer respond to existing treatments, incremental gains in speed and precision can translate into real therapeutic options. That medical potential is why many researchers argue that strengthening oversight, rather than restricting the technology outright, is the more realistic path forward.
What This Moment Actually Calls For
AI is not on the verge of mass producing bioweapons. What it is doing is reshaping how biological research begins, moving more early stage exploration into software and exposing weaknesses in safety systems that were built for an earlier era of biology.
What stands out is that these weaknesses are not being uncovered by outsiders or bad actors, but by researchers working within the system who are actively trying to stress test it. The same studies that reveal vulnerabilities are also producing practical fixes, from improved screening methods to clearer standards for oversight. That matters, because it shows that risk reduction does not have to come at the expense of medical progress.
For the public, the most useful response is informed attention rather than alarm. Understanding what these tools can and cannot do helps keep the conversation grounded in reality. It also helps ensure that safeguards evolve alongside the science, so that advances with real health benefits are developed responsibly instead of being slowed or sidelined by fear driven narratives.
