Radical Numerics, a San Francisco-based company, announced a new system called Omnii Defense that can identify natural and artificially designed pathogens. The tool runs on a standard laptop paired with a USB gene sequencer and does not require an internet connection, making it usable in remote field settings where connectivity is unreliable.
The founders, who previously contributed to the open-source DNA models Evo and Evo-2 during doctoral work at Stanford, created the company in 2025. Their earlier product, Omnii, a genome language model, incorporates DNA, RNA, protein and epigenetic data. In October the firm used Omnii to engineer bacteriophages that targeted a drug-resistant E. coli strain, a treatment administered under special FDA permission. A partnership with Tabula Bioscience was omitted from the announcement, prompting Harvard geneticist George Church, an adviser to both firms, to resign.
Omnii Defense evaluates a genetic sequence on several dimensions, estimating the probability that it is pathogenic, whether it may have been generated by AI, and its closest taxonomic relatives. The model then generates a threat score ranging from zero to one, where higher values indicate a greater perceived danger based on combined unnaturalness and functional predictions.
Earlier federal tools such as the Finding Engineering-Linked Indicators (Felix) rely on matching sequences against known organism libraries, which works when a malicious actor makes minor modifications to an existing pathogen. Omnii’s predictive approach differs; in a test where the Bundibugyo virus,currently linked to an Ebola outbreak,was removed from the training set, the model still flagged 33 percent of the virus’s sequences as pathogenic after only the species was omitted, and 22 percent after the entire genus was excluded, whereas non-AI methods gave no warning.
Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute of MIT and Harvard, emphasized the need for vigilance, stating, “We absolutely need to be vigilant about what AI is enabling, and if we can use technology to compete with it and hopefully get ahead of any kind of nefarious actors, that seems sensible.” She added that detecting a signal does not automatically translate into assessing real-world impact or engineered intent.
Access to Omnii Defense is presently limited to vetted external testers. Radical Numerics reports ongoing collaborations with national laboratories, government agencies and independent evaluators to validate the model’s performance and refine its threat-scoring methodology.