White House keeps finalized frontier AI framework under wraps

The White House says it hit its August 1 deadline to stand up a voluntary framework for evaluating the most advanced AI models, but is not showing anyone outside a small circle what the framework actually says. According to Axios, the finished document has not been released to industry or the public, and a White House official told the outlet that 'just because things are unclassified that doesn't mean we are going to broadcast them to everyone.'

The framework flows from a June 2 executive order that gave the government roughly 60 days to build a review regime for frontier models, with an option for up to a 30-day pre-release look at new systems. Anthropic, OpenAI and Google worked closely with the White House on the contours; the three labs jointly reviewed a draft in late July and submitted edits, per follow-up reporting, and the Office of the National Cyber Director is hosting the industry meeting to walk through next steps.

The part worth sitting with is who wrote it and who can read it. Standards meant to hold the largest labs accountable were drafted with heavy input from those same labs, and the cybersecurity benchmarks are being kept out of public view. The administration's line is that discretion is a security feature, not a bug, since the tests reportedly measure the cybersecurity capabilities of the most advanced US models and publishing them would defeat the point.

The honest caveat is that 'voluntary and unpublished' is hard to evaluate from the outside. Smaller labs that were not in the room have no way to know whether the definition of 'frontier' applies to them, civil society groups cannot challenge a benchmark they cannot read, and the reporting does not tell you what specifically triggers a review, what a lab commits to by opting in, or what happens if a model flunks a classified test.

For companies inside the tent, the upside is real. Early government access on a voluntary footing is a lot friendlier than the licensing regime Washington could have written, and the labs that helped shape the document get to operate against rules they helped calibrate. For everyone else, including open-source developers and the next tier of model builders, the shape of the deal is still an unknown drawn behind a closed door.