Radical Numerics Says Its Offline AI Can Flag AI-Designed Pathogens From a Laptop
Radical Numerics says its new Omnii Defense model can score natural and AI-engineered pathogen risks offline on a laptop. The figures are the company's own, and only vetted testers can use it.
Frontier AI labs keep warning that their models could help someone build a bioweapon. Radical Numerics, a San Francisco startup, now says it has built the other side of that fight, and it fits on a laptop.
On Wednesday, the company announced Omnii Defense, an edge model it says can flag both natural and AI-engineered pathogens using a laptop and a USB gene sequencer, with no internet connection required, Wired reported. The pitch is aimed at field teams with patchy connectivity. The catch is just as clear. Nobody outside a small circle can use it yet, and the headline numbers come from the company itself.
CEO Eric Nguyen frames the problem the way security people talk about malware. He told Wired he sees a coming arms race in biodefense similar to the one in cybersecurity, where attackers and defenders both lean on AI to get an edge.
"We felt that the biosecurity space was far behind on the defense side." Eric Nguyen, cofounder and CEO, Radical Numerics, to Wired
Nguyen and his cofounders helped build the open-source Evo and Evo-2 DNA models at Stanford and founded the company in 2025, according to Wired.
A score, not a verdict
According to Wired, the model makes several predictions about a genetic sequence, including the odds that it is pathogenic, the odds that it was AI-generated, and which taxonomic group it most closely resembles. It then rolls those into a single threat score between zero and one. Higher means more worrying, the company says.
Older tools such as the federally funded Felix program mostly look for matches against libraries of known organisms, Wired notes. Radical Numerics argues its model learns what a sequence does rather than what it spells. Christine Parthemore, the company's head of biodefense, called it a "step change."
The evidence offered so far is the company's own. In one experiment described to Wired, Radical Numerics removed Bundibugyo virus, the virus behind a current Ebola outbreak in East Africa, from the model's training data, first at the species level and then the entire genus. With the species removed, the company says the model still flagged 33 percent of the Bundibugyo sequences it was shown as pathogenic. With the whole genus gone, it flagged 22 percent. Parthemore said non-AI tools gave essentially no indication of risk. "It was as if Omnii was operating in 1976," she told Wired, referring to the year Ebola was discovered.
Interesting, but unaudited. And a model that catches roughly a fifth to a third of sequences from a virus it has never seen is a lead generator, not a gatekeeper.
Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome program at the Broad Institute of MIT and Harvard, told Wired the defensive logic makes sense.
"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." Bronwyn MacInnis, Broad Institute of MIT and Harvard, to Wired
She also drew the line that matters most. MacInnis cautioned that detecting a signal is still a big leap from determining whether a pathogen has real consequence or whether a threat was engineered at all. Wired adds its own reality check, noting that many scientists consider AI-made bioweapons a low-likelihood threat. This is insurance against a disputed risk, not a response to a proven attack.
There is some company baggage too. Wired reports that Harvard geneticist George Church resigned as an adviser last week after Radical Numerics left partner Tabula Bioscience out of an October research announcement.
For now, Omnii Defense is not a public product and not a diagnostic tool. Access is limited to vetted external testers, and the company says national laboratories, government agencies and other independent evaluators are assessing it. An October 20 company post had already described the restricted API and laptop setup. Those outside evaluations are the real test. Google DeepMind took a similar research-only line with AlphaGenome 2, and in biology the gap between a promising model and a trusted one is usually where the hard work starts.
