WASHINGTON — The White House has decided to exempt U.S. developers of open-weight artificial intelligence systems from a new voluntary federal security review process, narrowing the scope of an emerging framework aimed at assessing national security risks posed by the most advanced AI models, according to people familiar with the policy and administration guidance.
The policy, discussed by senior administration officials with leading AI companies this week, would apply only to certain closed, proprietary AI models that demonstrate cutting-edge capabilities in areas such as cybersecurity and computer hacking. Under the framework, developers of those frontier systems would be asked to voluntarily submit their models to the U.S. government for review before public release. Open-weight models, whose underlying model parameters are made publicly available for others to download and modify, would not be subject to those reviews at this stage.
The approach marks a more targeted oversight strategy as the administration seeks to address security concerns surrounding increasingly capable AI systems while avoiding broader restrictions that officials say could hamper U.S. innovation and competitiveness. The proposed reviews are voluntary rather than mandatory and are intended to focus on models considered most likely to present national security risks.
Representatives from major AI developers, including OpenAI, Anthropic, Google, Meta and Nvidia, met with White House officials to discuss the unpublished framework, according to people familiar with the meeting. Administration officials have not publicly released the detailed review criteria, and key elements of the process remain nonpublic.
The policy follows a June executive order directing federal agencies to strengthen AI security and cybersecurity while promoting U.S. leadership in artificial intelligence. The administration has also issued broader national security guidance calling for expanded adoption of advanced AI across defense and intelligence agencies while emphasizing testing, evaluation and security assurance for systems used in national security missions.
The exclusion of open-weight models has prompted debate among policymakers and AI governance advocates. Supporters argue that open models encourage innovation and broader research collaboration, while critics contend that increasingly capable publicly available systems could also present security risks if released without independent evaluation. Administration officials have indicated that the treatment of open-weight models could be revisited as the technology evolves.
The framework comes amid heightened scrutiny of advanced AI safety following recent disclosures by AI companies regarding cybersecurity incidents involving their systems, developments that have intensified discussions in Washington over how best to balance innovation with safeguards against misuse. For now, the administration's planned prerelease review process remains focused on a limited class of closed frontier AI models, while open-weight systems remain outside its scope.
