Radical Numerics unveils Omnii Defense, an offline model that scores DNA sequences for pathogen threat

Radical Numerics announced Omnii Defense on Wednesday, an edge model designed to flag natural and AI-engineered pathogens using a laptop and a USB gene sequencer. It can run without an internet connection, so researchers in the field with unreliable connectivity can use it, according to the company. Access is currently limited to vetted external testers. The report comes from Wired.

What Omnii Defense does

The model makes several predictions about a given sequence:

• the odds that the sequence is pathogenic
• the odds that it is AI-generated
• the taxonomic group it most closely resembles

It also assigns each genome a "threat score" between zero and one. The score combines the likelihood of unnaturalness with functional predictions. A higher score means a higher threat.

Parameter count, context window, pricing and a general release date have not been disclosed.

Background

The San Francisco-based company was founded in 2025 by researchers behind Evo and Evo-2, open-source DNA foundation models developed during their doctoral research at Stanford University. Omnii, the company's genome language model, is trained on biological information beyond DNA, including RNA, proteins and epigenetic signals. Omnii Defense is the company's latest tool built on that work.

In October, Radical announced it had used Omnii to create custom bacteriophages targeting a drug-resistant strain of E. coli. The cocktail was used to treat a sick patient with special permission from the Food and Drug Administration. Tabula Bioscience, another AI biotech startup, was also involved in that research, but Radical did not disclose the partnership in its announcement. Harvard geneticist George Church, an adviser to both companies, resigned from his role at Radical last week as a result.

How it differs from earlier tools

Earlier biodefense tools, such as the federally funded Finding Engineering-Linked Indicators (Felix), compare sequences of interest against libraries of known organisms. That works when a bad actor has made a few tweaks to an existing pathogen, but those tools are not designed to make predictions the way Omnii does. Radical says AI models can learn sequence patterns associated with biological function rather than relying on exact matches, so they can recognize similarities even when the genetic spelling has changed.

Company-reported test results

Radical removed Bundibugyo virus, which is currently driving an Ebola outbreak in East Africa, from the training data. It did so first at the species level and then for the whole genus. Christine Parthemore, Radical's head of biodefense, says this left Omnii operating as if it were 1976, the year Ebola was discovered.

• With only the species removed, the model still flagged 33 percent of the Bundibugyo sequences as pathogenic.
• With the whole genus removed, it flagged 22 percent of the viral sequences.
• Non-AI tools gave essentially no indication of potential risk, according to Parthemore.

These results come from the company. Radical says it is working with national laboratories, government agencies and other independent evaluators to assess the model's performance.

What this means

Omnii Defense targets sequences that library-matching tools can miss, and the 33 and 22 percent detection figures are the company's own evidence for that. Those numbers come from a single in-house experiment and have not yet been assessed by outside evaluators. Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, supports using technology to compete with AI-enabled threats. She cautions that detecting a signal is still a big leap from determining whether a pathogen has real consequence or whether a threat was engineered. Many scientists, the article notes, regard AI-made bioweapons as a low-likelihood threat. Until the independent evaluations Radical describes are complete, the model's real-world reliability remains open.