Trump's AI Order Seeks a 30-Day Look at Frontier Models

President Trump signed an executive order on June 2 that asks the country's leading AI companies to give the federal government an early look at their most powerful models before the public gets them. The order, titled Promoting Advanced Artificial Intelligence Innovation and Security, sets up a voluntary review window of up to 30 days so agencies can probe a model for what the text calls advanced cyber capabilities.

The mechanics are deliberately soft-touch. Participation is voluntary. Within 60 days, the National Security Agency and the National Institute of Standards and Technology are directed to build a classified benchmarking process that decides whether a given system counts as a covered frontier model. Companies that take part would submit those models for testing, and the government would help pick the trusted partners that get early access to them.

One line in the order matters more than the rest. It states plainly that nothing in the text authorises a mandatory licensing, preclearance, or permitting requirement for releasing new AI models. In other words, this is a request, not a gate. Reporting from NPR and The Register both stress that the framework leans on cooperation rather than compulsion.

A shorter leash than the first draft

This is not the order Washington expected a month ago. An earlier version, shelved in May, would have given the government up to 90 days to inspect advanced models. The White House pulled it over worries that a long review would slow American labs at exactly the moment they are racing Chinese rivals. The final text cuts the window to 30 days and wraps the whole thing in the language of innovation and national security rather than restriction. Scientific American read it as a notable shift in how the administration talks about the technology.

The cyber framing is not abstract. It arrives alongside programs like Anthropic's Project Glasswing, which puts offensive-security-grade models in front of critical-infrastructure operators. The order gives officials a hand in deciding which firms sit on the trusted-partner list for that kind of access.

The objection: who gets to be trusted

That last detail is where the criticism landed. If the government helps choose which companies get early access, it also gains a lever over them. Juan Londoño of the Cato Institute warned that the provision could open the door to weaponisation against companies that have any sort of conflict with the administration. Samir Jain of the Center for Democracy and Technology put it similarly, saying the order should not become a mechanism to punish companies for political or other arbitrary reasons.

For now the order asks rather than demands, and the benchmarks that will define a covered frontier model do not yet exist. The next 60 days, while the NSA and NIST write them, will say more about how heavy this touch really is. It lands in a year already crowded with rule-making, from the EU's move to ease its own AI Act to a patchwork of state laws taking effect this summer.