🧬Omnii Defense Flags AI-Made Bioweapons Offline, Hitting 33% on Bundibugyo

Your biosecurity tools just got an offline backup

Frontier AI labs have warned that their models could help people build dangerous bioweapons. Now a startup says it has a defense. Omnii Defense flags whether a genetic sequence looks like a threat, whether it came from nature or was engineered with AI, and it does this without an internet connection. If you work on biosecurity tooling, genomics pipelines, or public health surveillance, this matters because most detection tools depend on exact sequence matches. Omnii is built to catch signals without them, so it can spot a pathogen it has never seen. Researchers in the field with spotty connections can run it locally, which is the scenario where existing pipelines quietly fail. The headline numbers: the model flagged 33% of Bundibugyo sequences as pathogenic even with the species removed from training data, and 22% even with the whole genus removed. Non-AI tools gave essentially no indication of risk on the same sequences. Each genome gets a threat score between zero and one, plus predictions on odds of being pathogenic or AI-generated. The catch? Access is limited to vetted external testers. The startup is working with national laboratories, government agencies, and independent evaluators to check performance. Until those results are public, treat the 33% and 22% figures as the company's own benchmarks.