Trump's AI Order Wants a 30-Day Look Before Models Ship
A June 2 executive order asks frontier labs to hand models to the government a month before launch — and the model that triggered it was Anthropic's zero-day-finding Mythos.
The clearest sign that AI's capabilities have outrun its governance is when a single model release forces a new government policy. On June 2, the Trump administration signed an executive order asking frontier AI labs to give the government access to their most powerful models up to 30 days before public release. The order is careful, narrow, and — crucially — voluntary. But the thing that prompted it was not a hypothetical. It was a model that already exists, and that this magazine just wrote about: Anthropic's Claude Mythos, which demonstrated a superhuman ability to discover critical vulnerabilities in operating systems and browsers.
That lineage matters, because it explains the shape of the order. This isn't pre-emptive regulation aimed at a danger someone imagines might arrive. It's a reaction to a capability that was demonstrated in April and disclosed in numbers — tens of thousands of flaws found, a 27-year-old bug surfaced in OpenBSD for under $20,000. Washington watched a model do something that looked unmistakably like a national-security event, and moved.
What the order actually asks
The mechanism is an early-access window. AI companies are asked to provide "covered frontier models" to the government as much as 30 days ahead of public launch, so security agencies can evaluate them before they're loose in the world. The NSA makes the "covered frontier model" designation and runs the classified benchmarking that decides which models qualify. CISA and the Treasury co-develop the framework alongside it. The screening is explicitly aimed at one thing: advanced cyber capabilities — exactly the Mythos failure mode of a model that can find and weaponize software flaws faster than defenders can patch them.
Then comes the sentence that defines the whole order's character. It states, in plain language, that "nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement." There is no enforcement mechanism. A lab can decline. The government is asking, and it has written into the order that it cannot compel.
Voluntary is the entire story
That voluntariness is not a footnote — it's the policy. Reporting around the order describes a tougher May draft that was withdrawn, reportedly after concerns that hard mandates would slow American labs against Chinese competitors, and with input from AI-and-crypto czar David Sacks favoring a voluntary framework over a mandatory one. The administration landed on the lightest instrument that still puts government eyes on frontier models before launch: an invitation with a deadline, not a gate with a lock.
The bet underneath it is that the major labs will say yes anyway. And they plausibly will, because the incentives mostly point that way. A 30-day classified review is cheap insurance against being the company whose model gets blamed for the first state-scale autonomous cyber incident — and Anthropic, which built Mythos and then voluntarily withheld it through Project Glasswing, has already shown that at least one frontier lab will trade capability for caution when the dual-use risk is sharp enough. The order is, in effect, an attempt to turn one lab's restraint into a soft industry norm without the political fight of legislating it.
The two deadlines
The order runs on a clock. Agencies get 60 days to stand up the benchmarking process and the voluntary framework, and 30 days to push cybersecurity improvements across federal systems and critical infrastructure. The second timeline is the quiet tell: the government isn't only worried about screening future models, it's hardening its own defenses now, on the assumption that capabilities like Mythos either already exist outside the consortium or soon will. You don't sprint to patch your own infrastructure in 30 days because of a model that might ship next year. You do it because of one that shipped last quarter.
What's missing, and what to watch
The order leaves the most contested thing undefined on purpose: what counts as a "frontier model." There's no public compute threshold, no published parameter count, no transparent line. The criteria are classified, and the NSA director makes the final call. That's defensible on security grounds — you don't hand adversaries the exact spec of what trips the wire — but it also means the scope of the program is whatever the intelligence community decides it is, with no external check and no published standard for labs to plan around.
The deeper open question is whether voluntary survives contact with a refusal. The order works as long as labs cooperate. The first time a major model ships without going through the window — and nothing in the order can stop that — the administration faces a choice between accepting the gap and reaching for the mandatory authority it explicitly disclaimed. For now, this is the lightest-touch instrument that still does something: a government that watched an AI model find zero-days for pocket change, decided it had seen enough to act, and asked — politely, with a deadline, and without a stick — for a look before the next one ships.