US AI Safety: Open-Weight Models Excluded from Review

The White House has finalized a voluntary AI safety testing framework, explicitly excluding open-weight models from federal security reviews. This framework, implemented by CAISI under NIST, requires closed-source frontier models to undergo a 30-day voluntary early evaluation before release. This means that leading closed-source providers like OpenAI, Google, and Anthropic are subject to these rules. However, open-source models, including those from China like Kimi K3 and DeepSeek, do not need to be submitted to the U.S. government for testing. This framework effectively imposes constraints on America’s own proprietary vendors while offering a frictionless path for Chinese open-source models. This development highlights a significant shift in the approach to AI regulation, impacting the competitive landscape of AI development.