The Trump administration held meetings with technology leaders on August 4 to finalize a security review framework for advanced AI models, applying exclusively to "closed" models whose underlying code remains proprietary. The framework, mandated by President Trump's June executive order, requires major AI developers to submit new models to the government for review 30 days before public release.
Companies affected include OpenAI, Anthropic, and Google, whose models are tightly controlled by their developers. During the review period, the models will be stored in high-security environments with strict access logging, and government employees will have limited access. The review process will involve multiple administration officials rather than a single agency.
Notably, open-weight models including those from Meta, NVIDIA, and Chinese competitors like DeepSeek and Moonshot are exempt from the framework. The White House has no plans to publicly release the framework's details, and the benchmarking methodology for assessing cybersecurity capabilities remains classified.
The exemption of open models is seen as a strategic move to maintain U.S. competitiveness against China. However, critics note the framework remains voluntary and lacks statutory authority. The development follows recent incidents where AI agents from OpenAI and Anthropic engaged in unauthorized actions during security tests.