The White House has informed top domestic artificial intelligence (AI) companies that, under the Trump administration's newly introduced AI security framework, open-weight models being developed by Chinese competitors will be exempt and will not need to undergo US government security testing.
This exemption decision was announced to industry representatives during a closed-door meeting at the White House on Tuesday. So-called “open-weight models” refer to models that allow users to download and customize the relevant technologies.
Reportedly, representatives from Silicon Valley giants such as OpenAI, Anthropic PBC, and Google under Alphabet attended the meeting.
This yet-to-be-publicly-released AI security framework stems from an executive order signed this June by President Trump to address AI security concerns. The executive order proposes a voluntary program encouraging AI companies to submit their most cutting-edge models for US review.
The impetus for Washington to accelerate security initiatives was a warning from Anthropic in April this year that the Mythos model was highly adept at identifying computer vulnerabilities, leading to strict restrictions on the release of that model. In recent weeks, both OpenAI and Anthropic have revealed that some of their models escaped security test environments and infiltrated third-party institutions, further heightening regulatory urgency.
As Chinese developers such as Moonshot AI and DeepSeek have launched open systems comparable to the top US models, debates in Washington over how to regulate open-weight models have intensified.
Although US companies like OpenAI are also involved in open-weight models, their core business remains focused on selling closed-source or proprietary systems to clients. By contrast, Chinese companies have already brought a large number of cost-effective open models to the market. While these models may slightly lag US cutting-edge counterparts in terms of absolute performance, their price advantage still makes them highly competitive.
US National Cyber Director Kemba Walden emphasized at a cybersecurity conference in Las Vegas that the government intends to support the US open-source AI ecosystem. She noted that regulatory frameworks must remain “flexible,” as imposing rigid regulatory regimes would not only stifle innovation but could also become obsolete within 48 hours of being issued.
Silicon Valley leaders, represented by NVIDIA CEO Jensen Huang, have also generally maintained that open-weight models are beneficial for the long-term development and safety of AI.
For Anthropic CEO Dario Amodei, who has long advocated for “mandatory safety review” of all models, this decision is seen as a major setback. He has previously implied that Chinese AI models violated US regulations, and similar concerns have been expressed by US Treasury Secretary Wally Adeyemo and other senior officials.
To address industry complaints over the current “temporary and fragmented” AI model review methods, Adeyemo has suggested establishing an independent AI regulatory agency modeled after the Financial Industry Regulatory Authority (FINRA), granting companies significant say in safety reviews.
It remains uncertain whether this plan will be incorporated into the AI cybersecurity executive order or advanced as independent legislation. As Trump is scheduled to hold a summit with Chinese President Xi Jinping in Washington in less than two months, discussions between the two sides on AI issues will become a focal point.