White House finalizes secret 30-day pre-release AI review framework, exempting open-weight models

The White House finalized a framework requiring makers of closed, proprietary U.S. frontier AI models that demonstrate state-of-the-art cybersecurity and hacking capabilities to submit those models to the government for testing up to 30 days before release. The framework's contents have not been made public. Only closed, proprietary U.S. models meeting specific performance benchmarks are covered; open-weight models, including Chinese models such as Kimi K3, are exempt from the testing requirement. Nvidia and other open-weight model makers are initially exempt but may eventually be required to submit. One analysis characterized the framework as secret, mandatory in practice, and backward in design, arguing that labs releasing weights openly on platforms like HuggingFace face no scrutiny under it. The White House discussed the framework with executives from OpenAI and Anthropic. AI safety advocates criticized the secrecy of the process. One argument offered for keeping the guidelines private is that they could be strategically valuable for offensive cyber capabilities and defenses. HuggingFace CEO Clement Delangue's argument that model weights are like steel and not inherently dangerous drew criticism.