# Radical Numerics unveils Omnii Defense to flag AI-engineered pathogens offline > Source: > Published: 2026-11-11 17:05:33+00:00 # Radical Numerics unveils Omnii Defense to flag AI-engineered pathogens offline Photo: Steve A Johnson / Pexels The Menlo Park lab says its edge model can spot natural and AI-designed pathogens with a laptop and a USB gene sequencer, no internet required The worry that AI could help someone design a dangerous pathogen has hung over the field for a while. Now a startup says it has built the countermeasure. On Wednesday, Radical Numerics announced Omnii Defense, an edge model that flags both natural and AI-engineered pathogens. The hardware it needs fits in a backpack: a laptop and a USB gene sequencer. ## What Omnii Defense actually does The key feature is that it runs without an internet connection. That matters in outbreak zones, where reliable Wi-Fi is usually not the first thing anyone has. Omnii Defense belongs to the Omnii family led by Omnii-1, a multimodal genome language model from the Menlo Park-based AI lab. It works the way a chatbot learns English, except the text is DNA. Instead of predicting the next word, it learns the patterns in genetic sequences well enough to recognize what a pathogen looks like. According to the company’s research materials, Omnii-1 can identify pathogens whether they are natural, synthetic or AI-generated. It is offered through an API for large-scale screening and in an offline version built for field conditions. Traditional detection tools largely compare a sample against a library of known threats. A pathogen that has been mutated, or written from scratch by an AI, may not match anything in that library. ## The numbers behind the claim Radical Numerics benchmarked its approach against DIAMOND, an established bioinformatics tool. Omnii Defense posted a pathogen recall rate of **95.8%** at a 0.5 probability threshold. DIAMOND managed **78.6%** on the same measure. The gap matters most for what the company describes as paraphrased or divergent sequences. Those are genetic codes that do the same biological job as a known pathogen but are written differently. ### AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The model also flags metagenomic reads from viruses that were absent from its training data. Metagenomic reads are the jumbled fragments of genetic material pulled from a mixed sample, such as wastewater or a swab. The company says Omnii-1 catches reads from mutated viruses and novel strains that existing reference databases overlook. ## Who is behind it Radical Numerics was founded by Eric Nguyen and Michael Poli. Both worked on the Evo biological models, an earlier generation of AI systems trained on genetic data. The company raised a **$50 million** seed round in June 2026. Radical Numerics describes its mission as a dual mandate: advance biological AI while strengthening biodefense. Its stated focus spans both threat detection and therapeutic work. The company has expanded partnerships with national labs, including Lawrence Livermore. It has also brought on high-profile advisors with defense backgrounds, including Jill Hruby and Maj. Gen. Paul Friedrichs. ## The dual-use problem The company places safe and responsible AI development at the center of its strategy. Biological AI is a textbook dual-use technology. A model that understands how pathogens work could help design vaccines and therapies, or, in the wrong hands, help design something worse. Radical Numerics is betting that the best answer to AI-enabled threats is AI-enabled detection. ## What this means For public health agencies, the offline capability is the practical breakthrough. Surveillance that depends on shipping samples to a central lab or uploading data to the cloud is slow. A tool that works on a laptop in a remote clinic could shorten the time between a sample being taken and an alarm being raised. The API product hints at a second business. Large-scale screening could appeal to organizations that process huge volumes of genetic data and need to catch anything suspicious in the stream. Still, a benchmark is not a deployment. The 95.8% recall figure comes from the company’s own comparison against a single tool. Recall is also only half the story. A detector that catches everything but raises constant false alarms will quickly be ignored, so how often it flags harmless samples will matter to anyone deciding whether to rely on it. **Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).