Regulation

White House Frames Voluntary AI Review Around Closed-Source 'Frontier' Models

The White House’s voluntary framework for reviewing advanced AI focuses on closed-source models that pose national security risks, excluding open models from its scope.

White House Frames Voluntary AI Review Around Closed-Source 'Frontier' Models

The White House is excluding open‑source models from a voluntary framework intended to govern government review of advanced AI systems, according to people familiar with the matter who spoke to Axios.

Why this matters

The framework is meant to set how the Trump administration reviews advanced AI models before public release, but the White House does not plan to make the framework itself public.

What the framework contains

  • The draft framework reviewed on Tuesday classifies a "covered frontier model" as closed‑source, possessing state‑of‑the‑art capabilities and posing potential national security risks, multiple sources who attended White House meetings said.
  • The document, per those sources, does not provide a precise definition of what counts as state‑of‑the‑art or what threshold constitutes a national security risk.
  • Open models are explicitly excluded from the framework; the text states that nothing in the framework should be read as restricting open models after their release.

Practical elements of the pre‑release review

  • During a proposed 30‑day government review period before public release, employee access to the models would be limited.
  • Models would be housed in high‑security environments, and detailed logs would be kept of who accessed them.
  • The review process would involve multiple administration officials rather than being centralized in a single office or agency.

Broader context and access questions

  • Axios previously reported that the White House does not intend to publicly release the framework.
  • On Tuesday, the White House held staff‑level meetings with industry to discuss the framework; companies that were not invited remain unaware of its contents.
  • It is unclear which "trusted partners" would receive early access to advanced models under the framework, and whether any foreign governments would qualify as such partners.

Legal and policy context

  • The executive order explicitly states that the benchmarking process to assess advanced cyber capabilities of AI models will be classified.
  • The order does not require public release of the voluntary framework.

What’s next

  • During Tuesday’s meetings, companies were reportedly encouraged to share models that are as close to public release as possible, rather than early‑stage systems.
  • Sources said many companies with less advanced systems are likely to be left out of the framework.