For twelve years, OpenAI has been the company that puts frontier AI in everyone’s hands. GPT-2, GPT-3, GPT-4, ChatGPT. The whole narrative arc is about access. Sam Altman said it on stages, in blog posts, in congressional testimony. The mission is not to build the best AI for a few. It is to ensure the benefits of artificial general intelligence are broadly distributed.
And if you are a VP evaluating which AI vendor to build your roadmap around, that promise matters. It means the models get better, and you get them. Not eventually. Not after a review board. You get them.
On May 29, 2026, OpenAI launched GPT-Rosalind, its most capable biology model, under a program called Rosalind Biodefense. The announcement was careful. The language was warm. The access model was a locked door.
Rosalind is available to “vetted developers and U.S. government partners” through a “trusted access model that maintains security and accountability controls appropriate for advanced biological capabilities.”
Read that again. This is not a beta with a waitlist. This is not a staged rollout that reaches everyone in six weeks. This is a permanent gatekeeping architecture. The vetting is the product.
Three partners were named at launch. Lawrence Livermore National Laboratory. Johns Hopkins Applied Physics Laboratory. CEPI, the Coalition for Epidemic Preparedness Innovations. All three are government funded, government adjacent, or intergovernmental. Not one commercial enterprise. Not one startup. Not one university research lab without a defense affiliation.
OpenAI branded this “defensive acceleration.” The framing is that Rosalind exists to strengthen biosecurity. To help governments detect threats faster. To put frontier biology capabilities in the hands of the institutions best positioned to protect public health.
That framing is reasonable. It may even be correct. But it is a confession.
Because what OpenAI is saying, in language designed to sound like public service, is that its most capable domain model is too dangerous for broad access. That the founding principle of the company does not apply when the model is powerful enough to matter. That “open” was always a policy decision, not a principle. And policy decisions get revised when the capability curve crosses a threshold the founders did not anticipate when they chose the name.
The three partners will do important work. The biology model may genuinely advance biosecurity. None of that changes what the access model reveals.
The company that promised to distribute the future just decided who gets to hold it. Three institutions. All of them already had a seat at the table before OpenAI existed.
The rest of you can read the blog post.
👉 Read the agentic security news. Instantly analyze the threat vector, see if it applies to your setup, and find the gaps with our interactive playbook. All free.
👉 Follow me on LinkedIn | X | Substack for weekly analysis of real agent failures, control gaps, and what the frameworks are and are not catching.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.