The Biodefense Feedback Loop: Radical Numerics and the 'AI vs. AI' Arms Race
A new startup claims it can spot AI-engineered pathogens, but the narrative of an escalating biological arms race looks less like a security necessity and more like a convenient marketing cycle.
### Opinion: The Infinite Loop of AI Anxiety
There is a seductive logic currently dominating the venture capital landscape: the idea that the only cure for the dangers of artificial intelligence is more artificial intelligence. By framing the threat of AI-engineered bioweapons as an "arms race," the industry ensures a perpetual cycle of funding—building the weapon, then the shield, then updating the shield to fight the updated weapon.
***
### The 'Omnii Defense' Pitch
As Wired first reported, San Francisco-based startup Radical Numerics has entered this fray with the launch of Omnii Defense. The company claims this edge model can identify both natural and AI-engineered pathogens using only a laptop and a USB gene sequencer, notably functioning without an internet connection for use in remote field locations.
Founded in 2025 by a team that previously worked on the open-source DNA foundation models Evo and Evo-2 during their doctoral studies at Stanford University, Radical Numerics is positioning its technology as a necessary evolution in biosecurity. Eric Nguyen, the company's CEO and cofounder, explicitly compares the current state of biodefense to the cybersecurity sector, suggesting a coming arms race where both threat actors and defenders leverage AI to gain an advantage. Nguyen asserts that the biosecurity field has lagged significantly on the defensive side.
### Predictive Power vs. Catalog Matching
According to Wired, the primary distinction between Omnii Defense and previous tools is the move from matching to prediction. Earlier federally funded tools, such as Finding Engineering-Linked Indicators (known as Felix), operate by comparing genetic sequences against libraries of known organisms. While effective for spotting minor tweaks to existing pathogens, these tools rely on cataloged matches.
In contrast, Omnii Defense uses a genome language model that incorporates RNA, proteins, and epigenetic signals alongside DNA. This allows the model to learn patterns associated with biological function rather than relying on exact genetic spelling. The system assigns a "threat score" from zero to one, combining functional predictions with the likelihood that a sequence is unnatural or AI-generated.
To demonstrate this, Christine Parthemore, head of biodefense at Radical Numerics, described an experiment where the Bundibugyo virus—a driver of an Ebola outbreak in East Africa—was removed from the training data. Parthemore claims that even with the entire genus removed, the model flagged 22 percent of the viral sequences as pathogenic, whereas non-AI tools provided virtually no indication of risk.
### The Friction in the Narrative
Despite the technical claims, the rollout of Radical Numerics' technology has been marred by internal friction and external skepticism. Wired reports that Harvard geneticist George Church recently resigned as an adviser to Radical Numerics. The resignation followed a revelation that another AI biotech startup, Tabula Bioscience, was involved in research regarding custom bacteriophages used to treat a drug-resistant strain of E. coli—a project Radical had announced in October without disclosing Tabula's partnership.
Furthermore, the actual utility of these "threat scores" remains a point of contention. Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, notes that while using technology to counter nefarious actors is sensible, there is a significant gap between detecting a signal and determining if a pathogen has real-world consequences or was actually engineered.
Currently, access to Omnii Defense is restricted to vetted external testers. Radical Numerics states it is collaborating with government agencies, national laboratories, and independent evaluators to verify the model's performance.