Superpower Daily

Superpower Daily

Your reading space

Sign in
OpenAI Calls for Federal AI Safety Rules While Backing Four Cal...Anthropic Says It Removed Claude Accounts Tied to Nine Influenc...Pentagon Negotiates $5B Fluidstack Loan for AI Power and Coolin...TSMC Reports Record August Revenue as AI-Chip Demand Holds Stro...Enflame Shares Rise 206% in Shanghai Debut, Backing Its Next AI...Google Releases ToolGrad, Reporting a 99.8% Pass Rate for Tool-...OpenAI Calls for Federal AI Safety Rules While Backing Four Cal...Anthropic Says It Removed Claude Accounts Tied to Nine Influenc...Pentagon Negotiates $5B Fluidstack Loan for AI Power and Coolin...TSMC Reports Record August Revenue as AI-Chip Demand Holds Stro...Enflame Shares Rise 206% in Shanghai Debut, Backing Its Next AI...Google Releases ToolGrad, Reporting a 99.8% Pass Rate for Tool-...
Swipe

White House Has Not Released Its Finalized Frontier AI Testing Framework

The voluntary review path can give the government early access to advanced models, but its standards, participation and outcomes remain outside public view.

By 2 min read
White House Has Not Released Its Finalized Frontier AI Testing Framework
White House Has Not Released Its Finalized Frontier AI Testing Framework

Listen to this story

The audio brief

About 1:23
0:001:23
Read transcript
The White House has finished a framework for testing frontier AI models, but it still has not released the document publicly. As of September tenth, the government’s review process exists on paper while its standards and results remain out of view. The framework grew out of a June executive order. It gives companies a voluntary way to provide the government with access to their most advanced models up to thirty days before those models are released. The key word is voluntary: companies do not have to submit, and neither the administration nor the companies have to say whether a review took place. That means the public cannot see the tests used, identify every company that participated, or learn what the government concluded. A review could influence a model’s release without leaving a public record showing how. OpenAI has confirmed one submission. Its chief executive, Sam Altman, said the company provided its Astra model and found the process productive. But he did not disclose the tests Astra faced, the government’s findings, or whether the review led to any changes. The Center for Democracy and Technology is now urging the White House to publish the framework. That would allow outside observers to examine the rules and assess whether the voluntary channel is being used consistently. Until then, the central constraint is simple: early government access may exist, but participation and impact cannot be measured from the outside.

Story brief

3 key points

A White House framework for evaluating frontier AI models was completed by early August but remained unpublished as of September 10. The voluntary process lets companies provide models to the government up to 30 days before release, yet neither participation nor review results must be disclosed. OpenAI CEO Sam Altman confirmed that the company submitted its Astra model, without revealing the tests or findings. The...

  1. 01

    Companies can opt into federal review of advanced models before release; submission is not mandatory.

  2. 02

    The framework’s evaluation standards, participating companies, and review outcomes remain unavailable to the public.

  3. 03

    Sam Altman said OpenAI submitted Astra and found the process productive, but disclosed no test details or government conclusions.

The White House finalized a voluntary framework for testing frontier AI models by early August, but had not released it publicly as of September 10. The framework creates a federal review path for advanced systems while leaving the public unable to see the standards behind it or the results it produces.

The framework follows a June executive order that asked AI companies, including Anthropic and OpenAI, to voluntarily give the federal government access to their most advanced models up to 30 days before release. The order created an early-review option rather than a requirement for companies to submit their systems.

A completed process, still out of view

The administration is keeping the framework’s evaluation standards confidential. Neither the administration nor companies are required to say whether a review occurred or what it found, so a submission can be reviewed without creating a public record of its outcome.

What the framework does not require publicly

  • Disclosure of the standards used to assess a submitted model.
  • Confirmation that a company participated in a review.
  • Publication of a completed review’s outcome.

OpenAI identifies a submission

OpenAI CEO Sam Altman said the company had submitted its new Astra model for review and described the process as productive. His statement confirms that OpenAI used the channel, but it does not disclose the tests Astra faced, the government’s findings or any resulting changes.

That confidentiality is central to the framework’s practical limit. Early government access could allow a review before a model’s release, yet optional disclosure means outsiders cannot determine how widely companies use the process or how it affects release decisions.

A request for public scrutiny

The Center for Democracy and Technology has urged the White House to release the framework. Its request puts a concrete question before the administration: whether a voluntary system for reviewing advanced AI models should also expose its governing rules to public scrutiny.

Sources

  1. cbsnews.comWhite House framework for testing AI models remains hidden as concerns about threats mount

Loading discussion...

Related coverage

Continue reading