10/22/2026
🚨 Scientists just administered an AI-engineered virus to a human patient without peer review, clinical details, or independent verification.
San Francisco AI lab Radical Numerics (CEO Eric Nguyen) announced it has given a patient with a stubborn, antibiotic-resistant bacterial infection the first known AI-designed bacteriophage treatment.
Their genome language model, Omnii-1, generated tailored viral genomes: specifically modified T7-like phages with altered tail fibers.
These were designed to latch onto and kill a resistant E. coli strain that had already defeated both antibiotics and existing natural phages.
The company says the FDA authorized the trial.
Phages target bacteria only and do not infect human cells, so this is framed as experimental phage therapy, not a human pathogen.
Earlier this year, Stanford and the Arc Institute showed that AI can design fully functional bacteriophages from scratch (16 viable viruses out of 302 designs), published in Science.
Radical Numerics is claiming the next step: taking that capability into a living patient.
But major red flags remain.
The announcement is a self-published technical report, not a peer-reviewed study.
Radical Numerics has refused to release any clinical details, patient outcome, or the identities of the treating physicians, citing patient confidentiality and proprietary commercial interests.
Outside experts, including Harvard microbiologist Michael Baym, have expressed skepticism.
Without data, independent assessment is impossible.
Synthetic biology pioneer J. Craig Venter has previously warned that the same AI genome-design methods applied to dangerous pathogens would raise grave concerns.
A company just put an unvetted AI-designed virus into a human and will not show the results, the doctors, or the data.
That isn’t caution… that’s control.
If the treatment works, the public has no way to confirm it.
If it fails or spreads in unexpected ways, the same silence protects the company.
This is how experimental biology moves from lab to patient with almost no public record and almost no external check.