Donald Trump has signed an executive order asking the most advanced AI companies to submit their frontier models to government security review before public release. On the surface, this sounds like a significant concession to the national security wing of the administration. Under the hood, the details matter enormously — and the details are, at best, ambiguous.

The order asks companies to allow agencies including the NSA and the Defence Department to evaluate cutting-edge models up to 30 days before they are released. It stops short of mandating participation. It explicitly bars the creation of any new licensing or permitting regime for AI models. Companies that cooperate would be engaging voluntarily. The White House framing is that this is about collaboration, not control.

David Sacks, Trump’s former AI advisor, called the 30-day window a “game changer.” His argument: the shorter timeline means companies can engage with the government without meaningfully slowing down their release cycles. In the AI race, every day counts, he argues, and a month is a reasonable ask. Whether you find that persuasive probably depends on what you think the order is actually for.

Here’s what we know about the context. The order has had a rocky internal history. An earlier version reportedly included a 90-day review period. That version was scuttled in May after significant pushback from industry and parts of the administration friendly to the AI sector. The 30-day window is a compromise — short enough that companies can pretend it’s not a burden, long enough that the government gets meaningful access. Whether that’s the intent or the outcome depends on whether you think the government is being genuinely transactional or genuinely cynical.

The security rationale is real. Advanced AI systems have capabilities that, in the wrong hands or with insufficient guardrails, could cause serious harm. Letting national security experts review the most powerful models before they go live is not an unreasonable idea. The question is what “review” means in practice, what happens if reviewers find problems, and whether “voluntary” is a label that will survive contact with geopolitical reality.

Because here’s the thing about “voluntary” in this context. The top AI companies — OpenAI, Anthropic, Google, Meta — are all deeply intertwined with the US government already. They compete for government contracts, they appear before congressional committees, they have former national security officials on their boards. The question isn’t really whether they will share models with the NSA. They already do, to varying degrees. The question is whether this order creates a formal framework that: (a) normalises pre-release government access as standard practice (b) gives the administration leverage to demand more over time (c) sets a precedent that other countries — allies and adversaries alike — will cite as justification for their own review regimes

If you’re optimistic, this is a pragmatic middle ground between the heavy-handed licensing approach Biden considered and the anything-goes approach Trump’s early rhetoric suggested. If you’re not optimistic, “voluntary” is a word that will quietly disappear from the conversation once the order has been in force long enough to become institutionalised.

The bigger picture: the US government is now actively trying to understand and shape what happens at the frontier of AI development. That’s new. That’s significant. And the details of how this order is implemented — which models get reviewed, by whom, on what timeline, with what consequences — will tell you a lot more than the press release.