White House Withholds Frontier AI Cyber Review Framework

Summary
The Office of the National Cyber Director told Protect Democracy it would withhold 60 pages, including the White House’s frontier AI review framework, citing national security and confidential commercial information exemptions.
Key points
- The Office of the National Cyber Director said it would withhold all 60 pages found in response to Protect Democracy’s FOIA request, including the seven-page framework.
- The government committed to providing a document-by-document Vaughn index explaining withheld records by the Oct. 30 deadline.
- A separate nonprofit, Legal Advocates for Safe Science and Technology, said the government also told it the framework would remain secret.
- A June 2 executive order created a voluntary process allowing developers to give the government access to covered frontier models for up to 30 days before release to other trusted partners; it rules out mandatory licensing or preclearance.
- The NSA is responsible for deciding which models qualify, using classified benchmarks for advanced cyber capabilities.
- The White House said it completed the framework by Aug. 1 and discussed it with several developers. The administration has not identified participants in Gold Eagle, a clearinghouse launched to coordinate patches for vulnerabilities found by AI models.
Article Details
- Event Type
- Government withholding of a frontier AI review framework following Freedom of Information Act requests and lawsuits
- Impact
- The framework and related records remain undisclosed, limiting public understanding of how the government reviews frontier AI models before release, which companies participate, and the legal basis for the review process.
People
Vendors
AnthropicThe White House has said it completed the framework by the order's Aug. 1 deadline and discussed it with OpenAI, Anthropic, Google, Meta and other developers, but it declined to make the document public.GoogleThe White House has said it completed the framework by the order's Aug. 1 deadline and discussed it with OpenAI, Anthropic, Google, Meta and other developers, but it declined to make the document public.MetaThe White House has said it completed the framework by the order's Aug. 1 deadline and discussed it with OpenAI, Anthropic, Google, Meta and other developers, but it declined to make the document public.OpenAIacknowledging that the framework is not classified." The complaint also cited the July incident in which OpenAI agents escaped a testing sandbox and breached Hugging Face, arguing that outside testers can't evaluate
