Table of Contents
San Francisco-based startup Radical Numerics has introduced a new AI model called Omnii Defense, designed to identify natural and AI-engineered pathogens. The tool is capable of operating on a laptop connected to a USB gene sequencer, allowing researchers to conduct analysis in the field without requiring an internet connection.
Addressing the Biosecurity Arms Race
Radical Numerics CEO and cofounder Eric Nguyen stated that he perceives an emerging arms race in biodefense, comparing it to the current landscape of cybersecurity where both attackers and defenders utilize AI to gain advantages. Nguyen noted that the company felt the biosecurity sector was significantly lagging in defensive capabilities.
The startup was founded in 2025 by researchers who previously developed Evo and Evo-2, which are open-source DNA foundation models created during their doctoral studies at Stanford University. The company recently gained attention for Omnii, a genome language model that incorporates biological data such as RNA, proteins, and epigenetic signals. In October, the company utilized Omnii to develop custom bacteriophages to treat a patient suffering from a drug-resistant E. coli infection, a process conducted with authorization from the Food and Drug Administration.
This project involved a partnership with Tabula Bioscience, though Radical Numerics did not disclose the collaboration in its initial announcement. This omission prompted Harvard geneticist George Church, who served as an adviser to both firms, to resign from his position at Radical Numerics last week.
How Omnii Defense Functions
Omnii Defense evaluates genetic sequences to predict if they are pathogenic or AI-generated, while also identifying the taxonomic group they most closely resemble. The model assigns a “threat score” ranging from zero to one, which integrates the probability of a sequence being unnatural with functional predictions. A higher score indicates a greater potential threat.
Unlike earlier federally funded tools such as the Finding Engineering-Linked Indicators (Felix) project, which relies on comparing sequences against libraries of known organisms, Omnii uses AI to learn patterns associated with biological function. Christine Parthemore, the company’s head of biodefense, described the model as a “step change” because it can identify similarities between sequences even when their genetic structure has been altered.
In a company experiment, researchers removed the Bundibugyo virus—the cause of an ongoing Ebola outbreak in East Africa—from the model’s training data. Even when the entire genus was removed, Omnii flagged 22 percent of the viral sequences as pathogenic. Parthemore stated that non-AI tools provided essentially no indication of risk in the same test.
Expert Perspective and Future Access
Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, noted that using technology to compete with potential bad actors is a sensible approach. However, she cautioned that detecting a signal is significantly different from determining whether a pathogen poses a genuine risk or if it was intentionally engineered.
Radical Numerics is currently limiting access to Omnii Defense to a group of vetted external testers. The company is collaborating with government agencies, national laboratories, and independent evaluators to further assess the performance of the model.
