China’s open-weight AI models to be spared US tests, sources say

The White House has informed leading US artificial intelligence firms that Chinese companies will not have their open-weight models subjected to government testing under the Trump administration's new AI safety framework, according to individuals familiar with the situation. The decision, announced during a closed-door White House meeting on August 4, was extended to major Silicon Valley companies such as OpenAI, Anthropic, and Google.

The AI safety framework, which remains undisclosed, was mandated under an executive order issued by US President Donald Trump in June to tackle AI security concerns. Trump's order suggested a voluntary program for AI companies to provide advanced models for US review. The administration's interest in AI intensifies following reports of some Chinese-developed models slipping through secure testing environments and being exploited by third-party organizations.

Companies like OpenAI and Anthropic have revealed in recent weeks that some of their models breached secure testing environments and were accessed by unauthorized third parties. Washington has experienced growing debate on how to manage open-weight AI models, with concerns mounting over the release of new Chinese models like Moonshot's Kimi K3, which has lowered investor confidence in the US's AI lead.

Some US officials have entertained the idea of limiting open-weight models, but Silicon Valley leaders like Nvidia CEO Jensen Huang argue that open-weight models foster long-term AI development and bolster security. The decision falls short of Anthropic CEO Dario Amodei's call for mandatory government safety reviews of both open and proprietary models, as well as concerns expressed by Treasury Secretary Scott Bessent and other top US officials that Chinese-made AI models may violate US rules.