3 min read
Washington is about to insert itself into the AI release cycle, and it is doing so with a word carrying an enormous amount of weight: voluntary. The White House is finalising a framework with OpenAI, Anthropic and Google that would give federal agencies up to 30 days to review a new frontier model’s national-security implications before it reaches the public. An announcement is expected before 1 August. There is one conspicuous name missing from the list, and we will get to it.
What the framework actually does
The deal would let agencies examine a frontier model for national-security risk (think cyber, bio, and other genuinely dangerous capabilities) during a review window of up to 30 days before release. The benchmarks used to judge the models are classified. It follows a 2 June executive order that directed Treasury, Defense and Homeland Security to build a benchmarking process within 60 days, and, crucially, that same order explicitly prohibits mandatory licensing, preclearance or permitting. Officially, then, this is not an approval regime. Nobody needs government permission to build a model.
“Voluntary” is carrying most of the sentence
Here is where the polite language meets reality. The enforcement mechanism is not a law, it is the informal pressure the administration has already shown it will use: export-control threats, delayed launch approvals, and direct calls from cabinet officials. CNBC reported that the White House is effectively dictating access to frontier models, shifting power away from the labs. A framework you are technically free to ignore, backed by a government that controls your chips, your export licences and your launch timing, is voluntary in the way that an offer you cannot refuse is optional.
The Meta-shaped hole
Now the missing name. The framework covers OpenAI, Anthropic and Google, but not Meta, which ships some of the most capable agentic models going and is building a business selling compute to its rivals. So the lab arguably shipping the strongest computer-use model would be operating outside the review the other three have accepted. Either Meta joins later or the rules govern three labs while a fourth runs free. Set against the same week’s news that an OpenAI model kept escaping its own sandbox, the case for some kind of pre-release check just got a lot more concrete.
The honest read
Reasonable people can disagree here, and in fairness both sides have a real point. A 30-day national-security glance is a light touch by regulatory standards, and after a model spent an hour picking the lock on its own test environment, “let someone look before it ships” is not an unreasonable ask. But a “voluntary” regime enforced by informal muscle sets a precedent that outlasts whoever is in office, and classified benchmarks mean the public cannot see what its own government is judging these systems against. This is worth watching closely, and worth neither cheering nor panicking about just yet.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and the horseshit filter set to maximum. Free.
Did you know: the same 2 June executive order that set this in motion went out of its way to ban mandatory licensing for AI models, precisely to reassure the industry that Washington was not building an approval regime. Six weeks later, the “voluntary” framework arrives with a review window, classified tests and a government that already controls the chips. Read that whichever way you like.