Radical Numerics builds an AI model to detect biological threats

A model for field use

Leaders of leading AI labs have warned that their technologies could make it easier to develop dangerous biological weapons. Radical Numerics says it has built a defensive tool: Omnii Defense can run offline, including in places with unreliable connectivity.

Radical Numerics co-founder and CEO Eric Nguyen expects an arms race in biological security similar to what is happening in cybersecurity: attackers and defenders are using AI to gain an advantage. He says biological security is lagging behind.

Nguyen and his co-founders helped create Evo and Evo-2, open DNA foundation models they worked on during their doctoral studies at Stanford University. Radical Numerics was founded in 2025.

The San Francisco company previously introduced Omnii, a genome language model trained on different types of biological data, including DNA, RNA, proteins and epigenetic signals. In October, Radical said it had used Omnii to create bacteriophages—viruses that infect and kill specific bacteria. The phages were targeted at a drug-resistant strain of E. coli. A mixture of bacteriophages was used to treat a critically ill patient under special authorization from the U.S. Food and Drug Administration.

Another biotech startup, Tabula Bioscience, also took part in the study, but Radical did not mention the partnership in last month’s announcement. Afterward, Harvard geneticist George Church, who advised both companies, left his role at Radical last week.

How Omnii Defense assesses threats

The model analyzes a genetic sequence and estimates the likelihood that it is associated with a pathogen or was created with AI. It also identifies the taxonomic group of organisms that the sequence most closely resembles.

Omnii Defense assigns a genome a threat score from zero to one. It combines the likelihood of artificial origin with predictions about the sequence’s functions: the higher the score, the greater the estimated threat.

Earlier biodefense tools, including the federally funded Finding Engineering-Linked Indicators (Felix) system, compare genetic sequences of interest with libraries of known organisms. This kind of search for matches can help if an attacker has made slight changes to an existing pathogen. But these tools are not designed to make predictions the way Omnii is.

AI models can identify patterns associated with biological functions rather than search for exact sequence matches. This could let them recognize similarities even when the genetic “spelling” has changed. Radical Numerics’ head of biodefense, Christine Parthemore, called this approach a major change in capabilities.

The company tested the model on Bundibugyo virus, which is currently causing an Ebola outbreak in East Africa. Researchers excluded the virus from the training data first at the species level, then at the level of the entire genus. Parthemore said this was like asking Omnii to work in 1976, when Ebola was first identified.

When only the species was excluded from training, the model identified signs of pathogenicity in 33% of the Bundibugyo virus sequences it was given. When the entire genus was excluded, it flagged 22% of the viral sequences. Parthemore says conventional tools indicated virtually no potential risk in these experiments.

Bronwyn MacInnis, director of pathogen surveillance in the Infectious Disease and Microbiome Program at the Broad Institute of MIT and Harvard, considers it reasonable to use technology to counter dangers created by AI and stay ahead of attackers. But she cautions that detecting a signal does not establish how dangerous a pathogen is or whether it was artificially created.

Access remains limited

Omnii Defense is currently available only to vetted external testers. Radical Numerics is working with national laboratories, government agencies and independent experts to assess the model’s results.

Dataist's take: Omnii Defense could expand the ability to screen for unknown threats, but its results alone do not establish either a pathogen’s actual danger or its artificial origin. The company is still testing the model with external organizations, so its practical effectiveness has yet to be assessed.