Rosalind Biodefense

On May 29, OpenAI opened a program called Rosalind Biodefense: it pays for vetted biosecurity teams and a handful of US federal and allied public-health agencies to use GPT-Rosalind, a frontier life-sciences model that pharma firms already license but the public cannot touch. The model itself shipped in April. What is new is who gets in, and on whose dime.

The power to shave years off drug discovery and the choice to withhold it from most researchers are the same fact.

The work is concrete, not a slide deck. Lawrence Livermore is pairing the model with its supercomputers to design medical countermeasures; Johns Hopkins' Applied Physics Lab is folding it into protein-engineering screens; the vaccine coalition CEPI is aiming it at a specific Ebola species, Bundibugyo, under its pledge to stand up a vaccine within 100 days of a new threat. OpenAI's word for the strategy is 'defensive acceleration' — get a dangerous capability to the people guarding against outbreaks before it reaches anyone who would misuse it.

The reason this matters is the release shape, not the biology. The default for frontier AI has been broad availability; here the same model is rationed by a qualification gate, safety review, and hard refusals on known bioweapon red lines. OpenAI is not alone — Anthropic gated a security model to about fifty organisations in April, and DeepMind keeps its genomics model behind a research-only door. Three labs landed on the same architecture within weeks: the capability is real, so the question shifts from whether to ship it to whom you trust with it. Biology is where they judge the gap between defensive benefit and offensive harm to be narrowest.

The model carries the name of Rosalind Franklin, whose X-ray data cracked DNA's structure without her credit. A tool named for her is now withheld from most of the researchers who might want it — the power to compress drug discovery and the deliberate inaccessibility are the same decision.