White House AI testing framework excludes open models, lacks risk definitions

The White House’s voluntary AI testing framework, finalized after a June executive order requiring frontier AI developers to share models with the government pre-release, explicitly excludes open models from testing and cannot be used to restrict them post-release, according to Axios. The framework applies only to closed-source models deemed to have state-of-the-art capabilities and carry national security risks, but it does not define either term, creating ambiguity for developers seeking compliance guidance.

A 30-day grace period is specified for government review of covered models, though the framework’s voluntary nature means companies are not obligated to participate. Anthropic, OpenAI, and Google attended a White House briefing on the framework, but the administration has not released its details publicly.

The absence of definitions for 'state-of-the-art' and 'national security risk' leaves smaller providers uncertain about whether their models fall under the framework, while frontier labs have sought clarity to avoid triggering government restrictions.