Radical Numerics rolls out offline AI detector for engineered pathogens

Radical Numerics unveiled Omnii Defense on Wednesday, an edge model that flags both natural and AI-engineered pathogens. The Menlo Park lab says the whole kit fits in a backpack: a laptop and a USB gene sequencer. Crucially, it runs with no internet connection, which matters in outbreak zones where Wi-Fi is rarely the first thing available.

Omnii Defense sits in the Omnii family, led by Omnii-1, a multimodal genome language model that learns DNA the way a chatbot learns English. Rather than predicting the next word, it learns genetic patterns well enough to recognize what a pathogen looks like — natural, synthetic or AI-generated, per the company's research materials. It ships two ways: an API for large-scale screening and an offline build for field conditions.

The pitch targets a real gap. Traditional tools mostly compare a sample against a library of known threats, so a mutated pathogen or one written from scratch by an AI may match nothing. Radical Numerics benchmarked against DIAMOND, an established bioinformatics tool, and reported a pathogen recall rate of 95.8% at a 0.5 probability threshold, versus 78.6% for DIAMOND. The model also flags metagenomic reads — jumbled fragments pulled from mixed samples like wastewater or a swab — from viruses absent from its training data.

The team behind it has form. Founders Eric Nguyen and Michael Poli both worked on the Evo biological models, an earlier generation of AI systems trained on genetic data. Radical Numerics raised a $50 million seed round in June 2026, expanded partnerships with national labs including Lawrence Livermore, and brought on defense-heavy advisors Jill Hruby and Maj. Gen. Paul Friedrichs. The company frames its mission as a dual mandate: push biological AI forward while strengthening biodefense.

Caveats come attached to the headline number. The 95.8% recall figure is the company's own comparison against a single tool, and recall is only half the story — a detector that catches everything but cries wolf constantly gets ignored, so false-alarm rates will decide whether agencies trust it. What the offline build does change is speed: surveillance that depends on shipping samples to a central lab or uploading to the cloud is slow, while a laptop in a remote clinic could shrink the gap between swab and alarm.