The White House is asking OpenAI to slow roll the release of GPT 5.6 over safety concerns

Ronni Holmvig Strøm · 2026-06-26

OpenAI's next model, GPT-5.6, reportedly will not ship the way its predecessors did. Instead of a public launch, The Information reports (via TechCrunch) that the

OpenAI's next model, GPT-5.6, reportedly will not ship the way its predecessors did. Instead of a public launch, The Information reports (via TechCrunch) that the company plans to hand it to a small set of close partners first, because the Trump administration asked it to. At a staff meeting this week, Sam Altman reportedly said the government would be "approving access customer by customer" during a preview period, with a broader release to follow "a couple of weeks later" if the limited one goes well. Two agencies asked for the gated rollout: the Office of the National Cyber Director and the Office of Science and Technology Policy. OpenAI staff, per the report, "worked closely" with them on it.

Two Roads to the Same Playbook

Earlier this year Anthropic released its frontier cyber model, Claude Mythos, to a small group of partners through a program it called Project Glasswing. The argument was that the model was simply too capable to put in anyone's hands.

OpenAI is the company that built its reputation on the opposite instinct. Its whole deployment philosophy has been to ship to the public early and often, on the theory that the world adapts faster when it can see the thing. Now it is running the staged release play too, this time not as strategy or conscience but because the federal government asked.

A gated rollout for the most powerful models reached OpenAI from corporate differentiation on one side and federal pressure on the other, and it landed in the same place. On June 2, Trump signed an executive order asking AI companies to voluntarily submit models with advanced cyber capabilities for government testing up to 30 days before public release, with the government helping select the "trusted partners" who get early access. An administration that started "hands-off" now has a procedure. GPT-5.6 looks like its first real exercise.

The Gate Removes the Evidence

Staging a powerful cyber model is a reasonable thing to do. The capability in question is real. Large models have proven able to write malware, and the frontier worry is tools that can find and exploit software vulnerabilities faster than any human analyst, against systems that quietly run most of the economy. If a model can do that, customer-by-customer access during a preview is a sane way to learn what it does before everyone has it.

As TechCrunch notes about Mythos, "since these models remain closed to the public, it's difficult to tell just how much of a threat they really are." The closed preview is exactly what makes the danger claim impossible to check. "Too powerful to release" becomes unfalsifiable the moment release is the only way anyone outside the room could test it. The marketing reading and the safety reading of Glasswing stayed deadlocked for the same reason they always will: there was no evidence either side could point to that the other could examine.