Source time Edition date
A Startup Wants to Fight AI Bioweapons With AI
一家初创公司想用AI对抗AI生物武器
Worth reading for anyone tracking biosecurity, because Radical Numerics claims its offline model can spot both natural and AI-engineered pathogens, though the assertion still awaits outside verification.
Judgment
Vetted testers and the national labs now assessing Omnii Defense must decide whether a zero-to-one threat score is actionable, given that access is invitation-only and no outside evaluation has been published.
Basis
CEO Eric Nguyen frames biodefense as a coming arms race, saying the company felt the biosecurity space "was far behind on the defense side." Christine Parthemore, Radical Numerics' head of biodefense, calls Omnii a "step change" over match-based tools such as Felix: rather than comparing a sequence against libraries of cataloged organisms, the edge model learns patterns tied to biological function and assigns a threat score between zero and one combining unnaturalness with functional predictions. In one experiment the company removed Bundibugyo virus from training data; with only the species removed, the model still detected that 33 percent of the Bundibugyo sequences pushed to it were something pathogenic, and with the whole genus removed it flagged 22 percent of the viral sequences, while non-AI tools gave essentially no indication of risk. Bronwyn MacInnis of the Broad Institute says vigilance about what AI enables is sensible, but cautions that detecting a signal is a big leap from determining whether a pathogen has real consequence or whether a threat was engineered.
- Species removed33 percent
- Whole genus removed22 percent
Unclear The source gives no false-positive rate for the threat score, no independent benchmark, and no account of what the vetted testers or national laboratories have found so far.
Next Check whether Radical Numerics or a named national laboratory publishes Omnii Defense evaluation results outside the company's own experiments.