A laptop and a USB gene sequencer are the proposed front line for a new kind of biological threat detection. Radical Numerics says its Omnii Defense model can assess genetic sequences for signs of pathogenicity and possible AI generation, including when a sequence does not match a known organism exactly.
Looking for patterns beyond known matches
Traditional biodefense tools can compare a sequence against libraries of known organisms. The federally funded Finding Engineering-Linked Indicators, or Felix, is one example. This approach can help when a pathogen has been modified in ways that leave recognizable similarities to an existing organism.
Omnii Defense is designed to make predictions as well as comparisons. The company says it evaluates the odds that a sequence is pathogenic or AI-generated, and estimates which taxonomic group it most resembles. Its threat score runs from zero to one, combining a measure of possible unnaturalness with predictions about function. Higher scores indicate higher estimated threat.
The rationale is that AI models can learn patterns linked to biological function, rather than depending only on exact sequence matches. A changed genetic spelling may still preserve patterns the model has learned to associate with function. That could help identify sequences that conventional matching tools do not recognize as a close fit.
Testing what the model can recognize
Radical Numerics describes an experiment involving Bundibugyo virus, which is driving an Ebola outbreak in East Africa. The company removed the virus from the model’s training data, first at the species level and then at the level of its whole genus, to see whether Omnii could still detect signs of potential pathogenicity.
With the species removed, the model identified 33 percent of the Bundibugyo sequences presented to it as something pathogenic. When the entire genus was removed, it flagged 22 percent of the viral sequences. The company says non-AI tools gave essentially no indication of potential risk in this experiment.
Those results speak to a specific test, not a guarantee that the model will detect every unfamiliar biological threat. Christine Parthemore, Radical Numerics’ head of biodefense, called Omnii a “step change.” The experiment suggests the model can find some signal even when close examples are absent from its training data, but it does not by itself establish how useful that signal will be in every setting.
A field tool with limits
Radical Numerics says Omnii Defense can run on a laptop with a USB gene sequencer and work without an internet connection. That offline capability could make it usable by researchers in the field where connectivity is unreliable. The company’s stated goal is to help researchers and public health officials identify emerging biological threats.
Eric Nguyen, the company’s cofounder and CEO, sees biodefense entering an arms race similar to cybersecurity, where attackers and defenders both use AI in an effort to gain an advantage. He and his cofounders were part of the research team behind Evo and Evo-2, open-source DNA foundation models developed during their doctoral research at Stanford University. They founded Radical Numerics in 2025.
The company has also drawn attention for Omnii, a genome language model that incorporates biological information beyond DNA, including RNA, proteins, and epigenetic signals. In October, Radical said it used Omnii to create custom bacteriophages—viruses that infect and kill specific bacteria—to target a drug-resistant strain of E. coli. The cocktail was used to treat a sick patient with special permission from the Food and Drug Administration.
Another AI biotech startup, Tabula Bioscience, was involved in that research. Radical did not disclose the partnership in its announcement last month. George Church, a Harvard geneticist who advised both companies, resigned from his role at Radical last week after the omission.
Detection is only a first signal
Scientists have described AI-made bioweapons as a low-likelihood threat, even as the company argues for stronger defenses. Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, said vigilance about what AI enables is sensible.
MacInnis also cautioned that detecting a signal is a long way from determining whether a pathogen has real consequences or whether it was engineered. A model’s score can point researchers toward a sequence for further assessment, but the score is not itself a verdict about danger or origin.
Omnii Defense is currently available only to vetted external testers. Radical Numerics says it is working with national laboratories, government agencies, and independent evaluators to assess how the model performs. Those evaluations will help clarify what the tool can detect and how its predictions should inform decisions about biological risk.