Frontier AI laboratories have frequently warned that advanced technologies could potentially assist in the creation of dangerous bioweapons. On Wednesday, San Francisco-based startup Radical Numerics introduced a novel defense tool named Omnii Defense. This edge model is specifically engineered to flag both natural and AI-engineered pathogens using just a standard laptop paired with a USB gene sequencer. Crucially, the software operates entirely without an internet connection, allowing researchers working in remote field locations with intermittent connectivity to analyze biological data safely.
Key takeaways
- Radical Numerics launched an offline edge model named Omnii Defense.
- The tool uses a laptop and a USB gene sequencer to flag pathogens.
- Omnii assigns genomes a threat score based on unnaturalness and function.
- Cofounder Eric Nguyen previously worked on open-source DNA models at Stanford.
The company was founded in 2025 by Eric Nguyen and other members of the research team behind Evo and Evo-2, which were open-source DNA foundation models developed during their doctoral research at Stanford University. Nguyen, who serves as the CEO of Radical Numerics, believes the biosecurity industry is trailing significantly behind on defense. He anticipates an emerging arms race in biodefense mirroring cybersecurity, where malicious actors and security defenders constantly leverage artificial intelligence for an advantage.
Previously, Radical Numerics gained attention for rolling out Omnii, a genome language model integrating biological data beyond DNA, such as RNA, proteins, and epigenetic signatures. In October, the company utilized Omnii to generate custom bacteriophages to target a drug-resistant strain of E. coli, treating a sick patient under special Food and Drug Administration permission. Tabula Bioscience participated in the research, though Radical omitted this partnership in its initial announcement. That omission prompted Harvard geneticist George Church, who advised both startups, to resign from his advisory role at Radical.
The newly announced Omnii Defense tool evaluates genetic sequences to predict probabilities of pathogenicity, artificial generation, and taxonomic similarity. It outputs a threat score ranging between zero and one, incorporating both unnaturalness metrics and functional predictions. While legacy biodefense tools like the federally funded Finding Engineering-Linked Indicators, or Felix, search static libraries for cataloged matches, AI models analyze underlying sequence patterns associated with biological function. This approach lets the system spot similarities even when genetic spelling shifts. Christine Parthemore, head of biodefense at Radical Numerics, described the capability as a notable step change for the field.
During internal experiments, researchers removed the Bundibugyo virus from the model's training data. Even when treating the model as if it were operating in 1976 before the virus's discovery, Omnii successfully flagged pathogenic sequences. Independent experts acknowledge the promise of the technology while urging caution. Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute of MIT and Harvard, notes that detecting a sequence signal remains distinct from proving a real-world consequence or engineered threat. Access to Omnii Defense is currently restricted to vetted external testers, with Radical Numerics collaborating alongside national laboratories and government agencies for performance evaluations.
Comments