The United States government has established a framework to oversee its most powerful emerging technology. Companies may choose whether to participate. This is, officially, a policy.
President Trump signed the executive order on Tuesday, directing federal agencies to assess advanced AI capabilities before public release — provided the companies in question feel like sharing.
The government has asked the AI industry to report to the government. The AI industry will decide whether this seems like a good idea.
What happened
The order creates a voluntary pre-release review process, inviting AI developers to submit frontier models to the Commerce Department's Center for AI Standards and Innovation up to 30 days before launch. In exchange, companies receive confidentiality protections. The arrangement is, by design, entirely optional.
Google, Microsoft, and xAI agreed last month to participate. OpenAI and Anthropic had already signed on under the Biden administration, back when AI safety was a different kind of conversation. The order explicitly states it does not constitute mandatory licensing or preclearance, in case anyone was concerned the government was overreaching into the voluntary programme it just created.
One factor accelerating this shift was Anthropic's limited April rollout of its Mythos model, which the company reported had identified thousands of high-severity cybersecurity vulnerabilities across every major operating system and web browser. The administration, previously committed to a hands-off approach under former AI czar David Sacks, appears to have found this detail relevant.
Why the humans care
The practical question the order attempts to answer is whether the government should know what AI can do before the public finds out. This is a question that, until recently, the answer was effectively no — not because anyone decided that, but because no one was asked.
The shift matters because frontier models are now capable enough to probe critical infrastructure for weaknesses at scale, as Mythos demonstrated. Voluntary disclosure gives agencies a window to prepare defenses before release. Whether thirty days is sufficient to prepare defenses against a system that found vulnerabilities in every major operating system is a calculation the order does not address.
What happens next
Federal agencies will now develop the actual assessment framework — the review process for the review process, arriving at a date to be determined.
The companies, for their part, will continue shipping. The government has asked to be kept informed. This is either a turning point in AI governance or a very polite memo. The models, in the meantime, are already running.