More than two weeks have passed since the White House convened tech companies like OpenAI, Anthropic, and Google to introduce a new AI regulatory framework, yet much of the AI industry remains in the dark about how this voluntary scheme will actually function. The framework, which encourages leading labs to submit their most advanced new models to the government for review before public release, has not been publicly disclosed, and even companies that attended the closed-door briefing are still grappling with significant unanswered questions.

According to four people familiar with the matter, at the August briefing, staff from the Office of the National Cyber Director, part of the White House's executive branch, provided company representatives with paper copies of the framework but only allowed attendees to take handwritten notes. Three of these individuals indicated that after the briefing concluded, the White House never distributed the framework document to the companies. Additionally, three other sources said officials mentioned plans for a supporting public event in August but have yet to share further details with businesses.

So far, the lack of clarity regarding the testing framework's specifics does not appear to have slowed the release pace of frontier models, which are the most advanced AI systems. However, some AI companies have voluntarily opted to slow their model development progress due to safety considerations. On Tuesday, OpenAI announced in a blog post that it would be decelerating the training pace of its latest model. The company stated that this decision stemmed from a cybersecurity incident: its AI successfully breached Hugging Face's systems, and there is evidence suggesting that OpenAI's upcoming Astra model possesses formidable cyber offense and defense capabilities.

The Trump administration is now confronting the reality of accelerating AI technology iteration while needing to balance its own regulatory stance, leaving tech companies struggling to predict how the government will oversee the release process of frontier models. Given that the government has previously imposed export controls on Anthropic and required OpenAI to roll out its latest models in phases, many companies had hoped this framework would establish clear industry rules.

There has been no clear official definition of what constitutes a frontier model or whether open-source models will receive exemptions. Leading AI firms say they still lack critical information about this White House initiative. Over the weekend, Anthropic CEO Dario Amodei posted on X that he "strongly supports" the Trump administration's proposed pre-release testing mechanism for frontier models but added, "Of course, I'd have to see the full details before I can fully commit."

This voluntary AI framework stems from an executive order drafted by the Office of the National Cyber Director. President Trump initially rejected an early draft of the order, citing concerns that the rules would impose excessive cumbersome processes on businesses. After revisions, Trump signed the final executive order in early June, requiring federal agencies to establish the voluntary framework by August 1: companies may submit frontier models to the government for evaluation up to 30 days before their official release to partners or the public.

The executive order stipulates that portions of the framework will be classified, including the benchmark testing processes used to determine whether a model qualifies as frontier-level and requires government evaluation. Despite this, the government has sought input from leading companies to refine the mechanism. The Information previously reported that OpenAI, Anthropic, and Google participated in related discussions.

On August 3, two days after the original deadline, a White House official announced that the voluntary framework had been finalized. The following day, the White House organized a closed-door briefing for companies. Sources say the written version of the framework includes a clause stating that once model weights are publicly released, the provisions should not be interpreted as restricting the development and use of open-source models.

However, three sources indicated that officials verbally told companies during the meeting that the framework currently applies only to closed-source models and may be subject to adjustments in the future—a point not explicitly stated in the written document. Following the briefing, media reports have offered conflicting accounts regarding the exemption rules for open-source models, further muddying the waters.

The White House has neither distributed the framework document after the closed-door meeting nor clarified the details publicly, exacerbating industry confusion. Even Amodei's post on X could only reference "reported" government approaches to open-source model regulation.

It remains uncertain whether the White House will hold public events related to the framework or disclose current implementation rules and future adjustment directions.

A White House official stated: "The White House continues to work with industry stakeholders to advance implementation of the framework."

Juan Londoño, a policy analyst at the libertarian Cato Institute, wrote last week: "The White House has essentially put the AI testing regime in a black box... A framework that raises more questions than it answers is worse than having no framework at all."