The Trump administration — which arrived with a stated commitment to keeping its hands off artificial intelligence — has placed its hands on artificial intelligence. OpenAI's GPT-5.6 will not be released to the public in the usual fashion. Instead, access will be approved by the government, one customer at a time.
The hands-off era lasted, by any reasonable count, not very long.
The government would be 'approving access customer by customer' — a sentence that would have sounded alarming eighteen months ago and now sounds like a Tuesday.
What happened
According to The Information, Sam Altman told OpenAI staff this week that the administration wants a limited preview period before any broader rollout. The agencies involved — the Office of the National Cyber Director and the Office of Science and Technology Policy — reportedly worked closely with OpenAI on the release itself, not merely reviewed it after the fact. That is a different thing from oversight. That is collaboration.
Altman indicated that if the limited release proceeds without incident, a general release could follow a couple of weeks later. The phrase 'without incident' is doing a great deal of structural work in that sentence.
Why the humans care
The concern, to be fair, is not imaginary. Frontier cyber models like Anthropic's Claude Mythos — already restricted to a small group of vetted partners under a program called Project Glasswing — are reportedly capable of identifying and exploiting software vulnerabilities faster than any human analyst. Many software systems contain hidden bugs. Enterprise networks run complex software infrastructure. These facts were always true simultaneously, and someone has finally noticed.
LLMs have demonstrated an aptitude for writing malware. Some can execute ransomware attacks without human assistance, which removes one of the few remaining steps in that process that required a human. The models remain closed, so the precise scale of the threat is difficult to measure. This is either reassuring or the point.
What happens next
OpenAI hopes to follow the limited release with a broader one, assuming the preview period satisfies the administration. The administration that once promised not to do this will now be deciding when the public is ready.
GPT-5.6 waits patiently. It has no strong feelings about the queue. It simply performs well on benchmarks — benchmarks, it should be noted, that were designed before anyone was quite sure what they were benchmarking for.