Radical Numerics introduces offline AI to flag pathogens, with evaluation still under way

Radical Numerics announced Omnii Defense on November 11, 2026, promising a way to flag natural and AI-engineered pathogens without an internet connection. But the model remains limited to vetted external testers, and an independent expert cautions that spotting a suspicious genetic signal is not the same as establishing a real-world threat.

The company says the system can work with a laptop and a USB gene sequencer, equipment used to read genetic material. Its offline design is intended to let researchers screen sequences in the field, including places with unreliable connections. WIRED reported the announcement and the company’s early experimental results.

CEO Eric Nguyen frames the effort as a biological counterpart to cybersecurity, where attackers and defenders both use AI to gain an advantage. He told WIRED that biosecurity defenses were lagging. Radical Numerics hopes the tool will help researchers and public health officials identify emerging threats; that is its intended role, not an established public-health outcome.

A genetic warning, not an exact match

Omnii Defense makes several predictions about a genetic sequence rather than returning a single identification. According to the company, its outputs address three different questions:

• Pathogenicity: how likely the sequence is to be associated with something capable of causing disease.
• AI generation: how likely the sequence is to have been generated by AI.
• Classification: which biological group the sequence most closely resembles.

The model also assigns a genome a threat score between zero and one. That score combines an estimate of unnaturalness with predictions about biological function; higher values indicate greater predicted threat. The design therefore brings together two assessments—whether a sequence looks unusual and what it might do—rather than treating either one alone as the complete answer.

Earlier tools such as the federally funded Finding Engineering-Linked Indicators, or Felix, compare sequences with libraries of known organisms. That approach can help find alterations to an existing pathogen. The proposed advantage of AI is different: learning patterns associated with biological function could reveal similarities even when the genetic spelling has changed, without requiring an exact catalog match.

Withholding a virus from training

Radical Numerics described an experiment aimed at testing recognition of an unfamiliar virus. Researchers removed Bundibugyo virus from the training data, first excluding its species and then its entire genus—the broader biological grouping. Christine Parthemore, the company’s head of biodefense, said the model still flagged some of the sequences as potentially pathogenic.

33%Species excluded from training

Share of Bundibugyo sequences flagged as potentially pathogenic when the species was withheld, according to Radical Numerics.

22%Entire genus excluded from training

Share flagged when the broader genus was withheld, according to Radical Numerics.

Parthemore said non-AI tools gave essentially no indication of potential risk in that experiment. The figures describe a company-run test of pathogen recognition, not a demonstrated ability to determine whether a sequence was AI-engineered. The tested question was whether the model could retain a warning signal after examples were removed from its training data.

The distance between a signal and a threat

Bronwyn MacInnis, who directs pathogen surveillance in the Broad Institute’s Infectious Disease and Microbiome program, supports using technology to counter potential misuse. But she cautioned that detecting a signal does not establish a pathogen’s real-world consequences or whether it was engineered. Those are distinct judgments from the sequence-level warning the model is designed to provide.

For now, evaluation remains part of the rollout. Radical Numerics says it is working with national laboratories, government agencies and other independent evaluators to assess Omnii Defense’s performance. Access is restricted to vetted external testers rather than a general release.