AI
Startup Launches AI Tool to Detect Engineered Pathogens and Bioweapon Threats
Radical Numerics unveiled Omnii Defense, an offline-capable AI model that uses DNA sequencing to identify naturally occurring and artificially designed disease agents.
WIRED — AI
Leading artificial intelligence research institutions have cautioned that frontier AI systems pose risks for accelerating the development of biological weapons. Now a San Francisco biotech firm believes it has built a counteroffensive.
Radical Numerics introduced Omnii Defense this week—a compact AI model that runs on standard laptops equipped with USB gene sequencers to identify both naturally occurring and AI-designed infectious agents. Critically, the system operates independently of internet connectivity, making it deployable in remote research settings with unreliable networks.
Filling a Defense Gap#
The company's founder and leader, Eric Nguyen, draws a parallel to the current cybersecurity landscape, where both attackers and defenders are increasingly harnessing artificial intelligence to outmaneuver one another. He points out that biosecurity infrastructure lags significantly behind on the defensive side of this equation.
Nguyen and his research partners developed Omnii Defense after doctoral work at Stanford on Evo and Evo-2, foundational models trained on DNA sequences. The team established Radical Numerics in 2025.
How It Works#
The new defensive tool makes several simultaneous assessments of genetic sequences: estimating the probability of pathogenicity, determining if engineering occurred, and identifying the closest known taxonomic classification. These analyses combine into a single "threat score" between zero and one, where higher values indicate greater danger.
This approach differs from established biodefense systems like the federally supported Finding Engineering-Linked Indicators (Felix). Those older tools rely on comparing submitted sequences against known organism databases—essentially pattern-matching against a catalog. They excel when adversaries make minor modifications to known pathogens, but they cannot project predictions for novel variants.
AI-driven models, by contrast, extract deeper patterns in how genetic sequences correlate with biological function. This capability enables them to spot analogies across sequences that differ significantly at the molecular level.
Real-World Testing#
During validation experiments, Radical Numerics deliberately obscured training data for the Bundibugyo virus, which currently fuels an Ebola outbreak in East Africa. When only the species data was hidden, Omnii Defense still identified approximately one-third of Bundibugyo sequences as potentially dangerous. Removing the entire viral genus reduced detection to roughly 22 percent—far surpassing the results from non-AI alternatives, which failed to flag meaningful risk.

