A Startup Wants to Fight AI Bioweapons With AI

A new startup, Radical Numerics, claims to have developed a potential solution to combat AI-created bioweapons. The company unveiled its latest tool, Omnii Defense, which is designed to identify both naturally occurring and AI-engineered pathogens using a laptop and a USB gene sequencer. Notably, Omnii Defense can operate without an internet connection, allowing researchers with limited connectivity to utilize the technology.

While some experts argue that AI-designed bioweapons are an unlikely threat, Radical Numerics believes its innovation could assist researchers and public health officials in spotting emerging biological dangers. Co-founder and CEO Eric Nguyen describes the situation as a forthcoming arms race in biodefense, akin to the one currently occurring in cybersecurity, where both adversaries and defenders are utilizing AI to gain an advantage.

Nguyen and his co-founders were previously part of the research team behind open-source DNA foundation models called Evo and Evo-2, which they worked on during their doctoral studies at Stanford University. In 2025, they established Radical Numerics. The company recently garnered attention for its new genome language model called Omnii, which incorporates a variety of biological data sources, including RNA, proteins, and epigenetic signals.

In October, Radical Numerics showcased Omnii's capabilities by generating custom bacteriophages, viruses that target and destroy specific bacteria, including a drug-resistant strain of E. coli, used to treat a patient with proper FDA approval. However, the company's partnership with another AI biotech startup, Tabula Bioscience, was omitted from the announcement, leading to the resignation of Harvard geneticist George Church, who had advised both companies.

Omnii Defense's primary feature is its ability to generate multiple predictions about a genetic sequence, including its probability of being pathogenic or AI-generated and the taxonomic group it most closely resembles. It assigns a "threat score" between zero and one, combining the odds of unnaturalness with functional predictions.

Earlier biodefense tools, such as the federally funded Felix, compare genetic sequences against libraries of known organisms to identify potential threats. However, these tools lack the predictive capabilities of Omnii Defense. Omnii's advantage lies in its ability to recognize patterns in sequences associated with biological functions, even when minor changes occur in the genetic code.

This allows the model to identify similarities between sequences, even when their genetic structure has altered. Christine Parthemore, head of biodefense at Radical Numerics, describes Omnii as a "step change" in biodefense technology. In a test, the company removed the Bundibugyo virus, responsible for an ongoing Ebola outbreak in East Africa, from the model's training data.

Surprisingly, Omnii still detected that 33 percent of the Bundibugyo sequences pushed to it were pathogenic, even when the whole genus was removed. Non-AI tools, on the other hand, provided no indication of potential risk. Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute of MIT and Harvard, acknowledges the importance of remaining vigilant about the role AI plays in biodefense and leveraging technology to potentially outpace nefarious actors.

However, MacInnis cautions that detecting a signal is far from determining whether a pathogen poses a real threat or has been intentionally engineered. Currently, Omnii Defense is accessible only to selected external testers. Radical Numerics states that they are collaborating with national laboratories, government agencies, and independent evaluators to evaluate the model's performance.