Radical Numerics, which left stealth in June 2026 with a $50 million seed round, says it treated a patient’s multidrug-resistant bacterial infection with a bacteriophage cocktail — viruses that infect bacteria, not human cells — whose variants its AI model generated. The company posted a technical report on 20 October; The New York Times reported it that day. Nearly everything else is undisclosed.
- Key Point: No peer review, clinical outcome, authorisation date or FDA confirmation has appeared.
- Key Point: This was not a trial. The company cites an FDA single-patient expanded-access route at a physician’s request, but reports no registered study, public file number, phase 1 or ethics committee.
- Key Point: Omnii-1 altered selected regions of a known phage and ranked candidates; no variant count, sequences or parent comparison were released.
Compassionate use, not a trial
The company calls it a “physician-led, authorised compassionate use pathway” under “the oversight and safety requirements established by the relevant regulatory authorities,” without naming them. A spokesperson told the Daily Caller it was FDA single-patient expanded access, requested by a doctor. There is no registered trial, public protocol number, phase 1 or known ethics review; headlines saying “trial” are wrong.
Two points jar. The company invokes a US pathway, yet Marisa Azad, the infectious-disease specialist The New York Times presents as a collaborator, works at Ottawa Hospital in Canada, where the equivalent route differs. No source says where treatment occurred. The company also will not date the authorisation: “We are not sharing the specific authorisation date at this time,” its spokesperson said. The FDA declined to comment.
What Omnii-1 did — and what is hidden
The AI did less than headlines suggest, the company concedes. It did not design a viral genome from zero. It started from a known, well-characterised bacteriophage and made variants by editing selected regions “involved in particular biological functions,” while ranking candidates by predicted activity against bacteria. Before administration, the cocktail eliminated the patient’s resistant isolate in the laboratory.
The company says the phages are its own, citing “designs generated by Omnii-1” and a cocktail “containing Omnii-1-designed phage variants.” It releases no variant count, sequences or comparison with the parent virus — the evidence an outsider needs to judge novelty. That is the reservation raised by the only independent specialist quoted. Whether it worked is unknown: the company withholds the outcome, citing privacy. There are no efficacy data, no case-specific safety data beyond the generic claim that phages do not infect human cells and therapy is “generally safe,” and no follow-up period. The New York Times notes phage treatment lasts days or weeks while patients are monitored for months for relapse, so the announcement may predate the treatment’s end. The report’s most candid line is the company’s: “observations from a single patient cannot establish overall safety or efficacy.”
“If it works, it is incredible.” “My impulse is to be skeptical.” — Michael Baym, Harvard microbiologist unconnected to the work
Baym would wait for treatment to finish and for scrutiny of how new the viruses are. Chief executive Eric Nguyen says findings will be published through “peer-reviewed research.”
Biosafety concerns predate the case
In August, Stanford, the Arc Institute and the Broad Institute reported in Science on designing complete phage genomes with genomic language models, using ΦX174: of 285 synthesised and assembled genomes, 16 were viable, and some combinations beat resistant E. coli. The authors warned of “biosafety, biocontainment, and biosecurity considerations” and discussed safeguards centred on excluding sequences from training data.
In the same journal, Johns Hopkins researchers Thomas Inglesby and Moritz Hanke said generative genome design poses “urgent biosafety and biosecurity questions” and lacks governance. They, not the paper’s authors, argue that using such data to generate pathogen genomes infecting eukaryotes should not be pursued. Isaac Bogoch of Toronto General Hospital warned publicly that in the wrong hands this “could have catastrophic consequences.”
Data exclusions “successfully prevent the pretrained Evo 2 models from designing eukaryotic viruses.”
But that is the only documented safeguard, a per-lab choice, not a rule. The governance Johns Hopkins seeks does not exist.
The problem is presentational more than scientific, and presentation matters with a patient involved. A company introduced four months ago is claiming clinical history on its own blog while withholding the result, date and jurisdiction, and without saying whether its model generated the viruses or merely selected them.