President Trump signed an executive order on Tuesday asking America's frontier AI labs to hand over their most capable new models to federal testers for up to thirty days before public release. The order is voluntary. It explicitly forbids itself from creating any "mandatory governmental licensing, preclearance, or permitting requirement." And it lands in a White House that, as recently as last month, was not sure it wanted to sign anything at all.
The signing was quiet. Trump put pen to paper in private after postponing a planned ceremony with tech CEOs because, he told reporters, he "didn't like certain aspects of it." NPR reports the earlier draft gave the government ninety days to review covered models. The final text halves that to thirty. The shift is small in the abstract and significant in practice: thirty days is short enough that companies can hit competitive release windows, which is most of what their lobbyists were arguing for.
The order also instructs the Department of War, CISA, the Treasury and the National Security Agency to build out a benchmarking process for "covered frontier models", judged primarily on their "advanced cyber capabilities." That phrasing is a tell. The trigger for the whole exercise, as both NPR and CNBC report, was Anthropic's announcement in April that it was limiting the release of its Mythos Preview model because of its ability to find and exploit software vulnerabilities. The company's follow-up programme, Project Glasswing, brought senior administration officials including White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent into direct contact with the model's capabilities. Tuesday's order is, in effect, what happens when those conversations get written down.
The politics around it are messier than the document suggests. Trump's administration has spent its first months stripping back Biden-era AI rules and arguing that any guardrails would hand the lead to China. David Sacks, Elon Musk and Mark Zuckerberg reportedly lobbied against an earlier version of the order. At the same time, the Department of War has labelled Anthropic a "supply chain risk," prohibiting defence contractors from using its technology, and Anthropic is suing to reverse the designation. So the same agency now charged with assessing covered frontier models is actively in court with the company whose products forced the assessment to exist.
What the order does not do is at least as important as what it does. It creates no enforcement mechanism. It cannot compel a lab to participate. It does not require disclosure of training data, model weights, or evaluation results to anyone outside the federal partnership. Any regulation that bites, the order itself notes, would have to come from Congress, which is not close to passing one. The "AI cybersecurity clearinghouse" the order announces is a coordination body, not a regulator: it will deconflict scanning, route patches, and share information among voluntary participants.
That is roughly the shape of the bargain. The administration gets a framework it can point to without forfeiting its anti-regulation posture. The labs get a thirty-day review window they can plan around, and an official channel to negotiate what counts as a covered model. The harder question, the one neither side has answered, is what happens the next time a lab ships something like Mythos and the voluntary framework is the only thing standing between the model and the open internet.