# Radical Numerics' Omnii Defense flags AI-engineered pathogens on a laptop with a USB sequencer

> The offline model gives each genome a threat score and caught some Ebola-family sequences it was never trained on. Only vetted testers have it for now.

- URL: https://devisdead.ai/news/radical-numerics-omnii-defense-flags-ai-engineered-pathogens-on-a-laptop-with-a-
- Published: 2026-11-11T20:19:44Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: biosecurity, radical numerics, genomics, safety
- Originally reported by WIRED (https://www.wired.com/story/startup-fights-ai-bioweapons-with-ai/).

AI labs keep warning that their models could help someone build a bioweapon. Radical Numerics says it has a defense that also runs on AI, Emily Mullin reports for [WIRED](https://www.wired.com/story/startup-fights-ai-bioweapons-with-ai/). Its new edge model, Omnii Defense, spots pathogens, whether natural or designed with AI, with nothing more than a laptop and a gene sequencer that plugs in over USB, and it works offline.

## Who's behind it

The San Francisco startup was founded in 2025 by researchers behind Evo and Evo-2, open DNA foundation models from their Stanford PhD work. CEO Eric Nguyen expects a biodefense arms race like the one in cybersecurity, and says defense has fallen behind.

Its bigger Omnii model works across DNA, RNA, proteins and epigenetic signals. In October the company used it to design bacteriophages against a drug-resistant E. coli strain, used to treat a patient under special FDA permission. Radical didn't credit Tabula Bioscience, which also worked on that, and Harvard geneticist George Church, an adviser to both, resigned from Radical last week over the omission.

## How it works

For any sequence, Omnii Defense estimates whether it's pathogenic, whether it looks AI-generated and what group of organisms it most resembles. It combines those into a threat score from zero to one.

Older tools like the federally funded Felix look for matches against catalogs of known organisms, which helps if someone has tweaked a known pathogen. A model that learns patterns tied to function can, in theory, spot danger even when the genetic spelling has changed.

The company tested that by removing Bundibugyo virus, behind a current Ebola outbreak in East Africa, from training. With the species removed, the model still flagged 33% of Bundibugyo sequences as pathogenic. With the whole genus gone, 22%. Non-AI tools showed essentially nothing, according to head of biodefense Christine Parthemore.

## The caveats

Many scientists consider AI-made bioweapons unlikely. The Broad Institute's Bronwyn MacInnis supports fighting AI with AI, but notes that detecting a signal is a long way from knowing whether a pathogen matters or was engineered. Access is limited to vetted testers while national labs, government agencies and independent evaluators assess it.

**[Scoreboard](/scoreboard) update:** AI +1. An offline AI model spots some threats that sequence-matching tools miss.

Stay undead.