Radical Numerics unveils Omnii Defense, an offline model to flag AI-made pathogens

On Wednesday, the San Francisco-based startup Radical Numerics announced Omnii Defense, an edge model designed to flag natural and AI-engineered pathogens using a laptop and a USB gene sequencer. It can also run without an internet connection, so researchers in the field with spotty connections can use it. Frontier AI labs have warned that their technology could help develop dangerous bioweapons, and Radical Numerics presents this as a defense. Many scientists say AI-made bioweapons are a low-likelihood threat, but the company hopes the tool will help researchers and public health officials identify emerging biological threats.

Cofounder and CEO Eric Nguyen says he sees a coming arms race in biodefense, similar to the one in cybersecurity, where attackers and defenders both use AI to gain an edge. "We felt that the biosecurity space was far behind on the defense side," he says. Nguyen and his cofounders were on the research team behind Evo and Evo-2, open-source DNA foundation models they built during doctoral research at Stanford University. They founded Radical Numerics in 2025.

Omnii Defense builds on Omnii, a genome language model the company rolled out recently. Omnii is trained on more than DNA: it also uses RNA, proteins and epigenetic signals. In October, Radical announced that it had used Omnii to create custom bacteriophages, viruses that infect and kill specific bacteria, aimed at a drug-resistant strain of E. coli. Its cocktail was used to treat a sick patient, with special permission from the Food and Drug Administration.

Omnii Defense makes several predictions about a sequence, including the odds that it is pathogenic or AI-generated and which taxonomic group it most closely resembles. It gives a genome a "threat score" between zero and one that combines the likelihood of unnaturalness with functional predictions; the higher the score, the higher the threat.

The article contrasts this with earlier biodefense tools such as the federally funded Finding Engineering-Linked Indicators, or Felix. Those compare sequences of interest against libraries of known organisms, essentially searching for cataloged matches. That helps when a bad actor has made a few tweaks to an existing pathogen, but such tools are not designed to make predictions the way Omnii is. AI models can learn patterns in sequences associated with biological function instead of relying on exact matches, which could let them recognize similarities even when the genetic spelling has changed. Christine Parthemore, Radical's head of biodefense, calls Omnii a "step change."

The company's evidence is one experiment. It removed Bundibugyo virus, which is currently driving an Ebola outbreak in East Africa, from the training data, first at the species level and then the whole genus. "It was as if Omnii was operating in 1976," Parthemore says, referring to the year Ebola was discovered. With only the species removed, the model still detected that 33 percent of the Bundibugyo sequences pushed to it were something pathogenic. With the whole genus removed, it flagged 22 percent of the viral sequences. Non-AI tools gave essentially no indication of potential risk, she says.

Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, backs the general idea: "We absolutely need to be vigilant about what AI is enabling, and if we can use technology to compete with it and hopefully get ahead of any kind of nefarious actors, that seems sensible." She cautions, however, that detecting a signal is still a big leap from determining whether a pathogen has real consequence or whether a threat was engineered.

There is also a governance wrinkle. Another AI biotech startup, Tabula Bioscience, was involved in the October bacteriophage research, but Radical did not disclose the partnership in its announcement. The omission led Harvard geneticist George Church, an adviser to both companies, to resign from his role at Radical last week.

Access to Omnii Defense is currently limited to vetted external testers. Radical Numerics says it is working with national laboratories, government agencies and other independent evaluators to assess the model's performance.

Key facts

Radical Numerics announced Omnii Defense on Wednesday: an edge model meant to flag natural and AI-engineered pathogens using a laptop and a USB gene sequencer, able to run without an internet connection.
The model outputs a "threat score" between zero and one, plus odds that a sequence is pathogenic or AI-generated and the taxonomic group it most resembles.
In a company experiment with Bundibugyo virus removed from training data, Omnii still detected 33 percent of Bundibugyo sequences as pathogenic with only the species removed, and flagged 22 percent of viral sequences with the whole genus removed.
Broad Institute's Bronwyn MacInnis calls the approach sensible but says detecting a signal is a big leap from judging whether a pathogen matters or was engineered.
George Church, an adviser to both Radical and Tabula Bioscience, resigned from Radical last week after Radical did not disclose the Tabula partnership in its October phage announcement; access to Omnii Defense is limited to vetted testers.

Why it matters

Frontier AI labs have warned that their models could help build bioweapons, and this is a startup's attempt at the defensive side. The pitch is that a model trained on biological function can recognize a threat whose genetic spelling has changed, where older tools such as Felix only search libraries of known organisms for cataloged matches. Nguyen expects a biodefense arms race resembling the one in cybersecurity. The article also notes that many scientists regard AI-made bioweapons as a low-likelihood threat, so the problem itself is contested.

Who it affects

The stated users are researchers and public health officials, including researchers in the field with unreliable connections, since the model can run offline on a laptop with a USB gene sequencer. For now only vetted external testers have access. Radical says it is working with national laboratories, government agencies and other independent evaluators. The article names none of them.

How to use it

It is not open to the public yet. Access is limited to vetted external testers, and the article gives no release date or pricing. In use, a sequence from a USB gene sequencer goes into the model on a laptop, and the output includes the odds of the sequence being pathogenic or AI-generated, the closest taxonomic group, and a threat score from zero to one, where a higher score means a higher threat.

How solid is it

Thinly supported so far. The 33 and 22 percent figures come from Radical's own experiment, as described by its head of biodefense, Christine Parthemore; no independent verification is reported. The article does not say how many Bundibugyo sequences were tested, and gives no accuracy, false-positive rate or threshold for what counts as a high threat score. The claim that non-AI tools gave essentially no indication of risk is also Parthemore's. An outside expert, Bronwyn MacInnis of the Broad Institute, supports the general approach without endorsing the results.

Risks and caveats

MacInnis cautions that detecting a signal is a long way from knowing whether a pathogen has real consequence or whether a threat was engineered. The model's best result in the experiment flagged 33 percent of the target virus sequences, and 22 percent at genus level. Separately, Radical did not disclose the Tabula Bioscience partnership in its October bacteriophage announcement, and the omission led adviser George Church to resign from his role at Radical last week. The article does not state Church's reasons beyond the omission, and gives no comment from Church or Tabula.

“We felt that the biosecurity space was far behind on the defense side”

— Eric Nguyen, cofounder and CEO of Radical Numerics