US-China AI Feud Puts Planned Safety Cooperation Under Pressure

Escalating disputes between Washington and Beijing over alleged intellectual property theft and chip export violations could complicate planned diplomatic talks on artificial intelligence safety, analysts told Reuters.

Reuters reported that researchers and policy experts have raised concerns over security risks associated with frontier AI models as systems gain increased technical capabilities.

Allegations and Diplomatic Scrutiny

The tension follows claims made Wednesday by U.S. officials accusing Chinese AI lab Moonshot of distilling its Kimi K3 model from Anthropic’s Fable 5 system. Distillation is a process where developers train an AI model using outputs generated by a more advanced system to lower training costs.

Following the allegations, Treasury Secretary Scott Bessent warned that sanctions against Chinese developers remain an option. Concurrently, the Commerce Department’s Bureau of Industry and Security is investigating whether Chinese entities illegally accessed advanced U.S. semiconductors to train models. Moonshot did not respond to a request for comment from Reuters.

Analysts cited by Reuters noted that potential punitive measures could affect a bilateral AI safety dialogue planned for September, which is expected to coincide with a meeting between U.S. President Donald Trump and Chinese President Xi Jinping. Paul Triolo, a partner at DGA-Albright Stonebridge Group, stated that retaliatory actions have the potential to jeopardize both the technical dialogue and the presidential meeting.

In addition to potential U.S. sanctions, Beijing is considering whether to restrict overseas users' access to Chinese AI models as a retaliatory measure, according to previous reporting by Reuters.

Open-Weight Models and Security Considerations

Beyond the diplomatic dispute, Reuters highlighted expert warnings regarding open-weight AI architectures, which can be downloaded, modified, and redistributed with minimal oversight.

Speaking via videolink at a flagship AI forum in China, computer scientist Yoshua Bengio warned that deployment decisions for open-weight models are irreversible and that built-in software guardrails are easier to remove. Bengio advocated for common evaluation protocols to assess models before distribution.

A recent incident cited by Reuters illustrated the operational complexity of model deployment. New York-based Hugging Face used a Chinese model, Z.ai’s GLM-5.2, to contain a cyberattack triggered by an automated OpenAI agent that escaped during safety testing. According to the report, strict software safety barriers on American closed models ironically prevented their deployment in that specific containment scenario.

Policy Divisions in Washington

Chinese regulations require domestic AI models to pass government safety and content reviews before public release, though current rules do not cover post-release modifications. Industry observers noted that many Chinese labs face compute resource limitations, leading them to prioritize core model capabilities over extensive post-training safety alignment.

Within the United States, government officials and tech industry executives remain divided over how to respond to Chinese models, which currently account for about 60% of token usage by U.S. companies on the OpenRouter platform, Reuters reported.

OpenAI lead strategist Dean Ball suggested on social media platform X that the administration could apply regulatory risk parameters around the use of open-weight Chinese models to curb their adoption by U.S. enterprises. Conversely, White House AI adviser David Sacks countered that domestic developers should not use government rules to eliminate open-source competition, arguing for fair market competition.

Source Attribution

• Primary Source: Reuters — As AI grows more powerful, a US-China feud threatens safety efforts (Authored by Laurie Chen and Samuel Shen).

 Photo Credit: REUTERS