Startup builds offline AI tool to flag engineered pathogens

Radical Numerics announced on Wednesday an edge model called Omnii Defense designed to flag natural and AI-engineered pathogens using a laptop and a USB gene sequencer, with no internet connection required. The San Francisco startup assigns each sequence a threat score between zero and one that combines likelihood of unnaturalness with functional predictions. For researchers and public health officials, the pitch is faster field screening of emerging biological threats.

How Omnii Defense scores pathogen risk

Omnii Defense makes multiple predictions about a sequence, including the odds of it being pathogenic or AI-generated and what taxonomic group it most closely resembles. The higher the threat score, the higher the assessed threat. Cofounder and chief executive Eric Nguyen said the biosecurity field was far behind on defense, describing a coming arms race similar to cybersecurity where attackers and defenders both use AI. The founders previously worked on Evo and Evo-2, open-source DNA foundation models from their doctoral research at Stanford University, and founded Radical Numerics in 2025.

The company built Omnii Defense on top of Omnii, a genome language model trained on DNA, RNA, proteins and epigenetic signals. In October, Radical said it used Omnii to create custom bacteriophages targeting a drug-resistant strain of E. coli, with a cocktail used to treat a patient under special permission from the Food and Drug Administration. Another startup, Tabula Bioscience, was also involved in that work but was not disclosed in the announcement, leading Harvard geneticist George Church, an adviser to both companies, to resign from Radical last week. Earlier federally funded tools such as Finding Engineering-Linked Indicators, or Felix, compare sequences against libraries of known organisms, which helps when a known pathogen was only slightly altered but does not make functional predictions in the same way.

What field screening means for work and health

For field researchers, clinicians and lab staff, offline screening could change where pathogen checks happen and how quickly suspicious samples are triaged. Head of biodefense Christine Parthemore called the model a step change because AI models can learn patterns linked to biological function rather than relying on exact matches. In a company test, Bundibugyo virus, now driving an Ebola outbreak in East Africa, was removed from training data to simulate discovery conditions in 1976, the year Ebola was discovered. With the species removed, the model flagged 33 percent of Bundibugyo sequences as pathogenic, and with the whole genus removed it flagged 22 percent, while non-AI tools gave essentially no indication of risk. Access is currently limited to vetted external testers, with national laboratories, government agencies and independent evaluators assessing performance.

Detection is not proof of danger or engineering, and many scientists view AI-made bioweapons as a low-likelihood threat. Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, said vigilance about AI-enabled risks is sensible but cautioned that a signal is far from showing real consequence or deliberate design. The test results are company-reported and have not been independently confirmed in the source, and the tool does not replace laboratory confirmation or public-health investigation. For most readers outside biodefense and infectious-disease work, there is no direct effect on jobs or daily health routines.