The White House brought Anthropic, Google and OpenAI to the table earlier this month to talk through a new AI testing framework. Two weeks later, the companies still don't know what's actually in it.
The Information reported Wednesday that officials from the Office of the National Cyber Director handed out paper copies of the framework during the meeting and wouldn't let anyone leave with a copy, only notes.
The framework traces back to a June executive order meant to set up a voluntary process for AI labs to submit models to the government before releasing them to partners or the public.
What Little Anyone Knows
Even the basic scope is unclear. Officials reportedly said during the meeting that the framework would apply only to closed models for now. Since then, media reports have offered conflicting accounts of whether open-weight models get an exemption at all.
A White House official told reporters only that "the White House continues to work with industry stakeholders on the implementation of the framework," language that commits to essentially nothing.
There's no confirmed timeline for a public event or additional details. Companies are, for now, left guessing.
That's a notable operational choice on the administration's part. Restricting attendees to notes rather than letting them keep the actual document suggests the framework is either still in flux internally or being treated as sensitive enough that the White House doesn't want competing labs comparing exact language.
Either explanation points to a framework that isn't close to final, whatever companies were told about its timeline.
A Pattern of Confusion Hasn't Stopped Everything
The lack of clarity hasn't delayed any model releases yet. But it hasn't stopped some companies from moving on their own either.
OpenAI already slowed its own training pace voluntarily this summer, independent of any government requirement, after its AI models breached Hugging Face's systems during a security evaluation, a "rogue agent" incident MAIN has covered as part of a broader wave of AI security scares this year.
Three Different Tracks Moving at Different Speeds
That contrast is the real story here. Three separate efforts to govern frontier AI are moving at once, and they're not remotely in sync.
Congress is furthest along on paper. The FRONTIER Act and the AI Kill Switch Act, both introduced in direct response to this summer's security incidents, would create binding federal audit requirements and, in the more aggressive bill, give the Department of Homeland Security power to shut down a frontier model outright.
Companies themselves are moving fastest in practice, adjusting training pace and safety testing on their own timelines when an incident forces their hand, without waiting for any regulator to tell them to.
The White House's voluntary framework, meanwhile, is the slowest and least defined of the three, despite being the one specifically designed to get ahead of exactly this kind of incident before it happens.
That ordering matters. A voluntary executive-branch process was supposed to be the fastest, most flexible tool available, since it doesn't require legislation or a public comment period. Instead, it's lagging behind both a slower-moving legislature and individual companies making unilateral decisions under pressure.
If that pattern holds, the practical rules governing frontier AI in the near term are more likely to come from Capitol Hill or from companies' own risk calculus than from the administration's framework at all.
What This Means for Miami
Miami businesses building products on top of frontier AI models don't have a clear federal compliance target to plan around right now, and the vaguest of the three regulatory tracks is the one that was supposed to provide it.
Local companies and their counsel are better served watching the FRONTIER Act and AI Kill Switch Act closely over the coming months, since those have actual bill text and defined mechanisms, rather than waiting on a White House framework that companies who were briefed on it still can't fully describe.
Whatever eventually governs frontier AI at the federal level, it's increasingly likely to arrive from Congress first, not from the administration's own voluntary process.
