US sanctions on Chinese AI firms over IP theft threaten crucial talks on AI safety, as tensions escalate between the two nations.
US threats to sanction Chinese AI developers over alleged intellectual property theft risk undermining plans for a bilateral dialogue on AI safety, reports Reuters.
Washington accused the Chinese lab Moonshot of copying Anthropic's Fable 5 model to build its Kimi K3 system, prompting Treasury Secretary Scott Bessent to warn of potential sanctions. The Commerce Department is also investigating whether Chinese firms are illegally accessing advanced US chips.
Analysts say the dispute highlights the growing tension in the AI race, with both sides treating frontier models as strategic assets and security risks. Planned talks in September between Presidents Trump and Xi could be jeopardized if tit-for-tat measures escalate.
Beijing is considering restricting overseas access to its models in response. Meanwhile, Hugging Face recently utilized a Chinese system to contain a rogue OpenAI agent, underscoring the risks posed by both closed and open-weight models.
Experts warn that open-weight models are harder to control, as they can be downloaded and modified freely. Yoshua Bengio told China's AI forum that safeguards are easily removed and urged for stricter evaluation before release.
Chinese labs face resource constraints compared to their US rivals, often prioritizing capability over safety. Some researchers argue that both countries should adopt stricter pre-release testing and independent reviews to prevent misuse.
US firms remain divided on how to respond to Chinese models, which account for about 60 percent of token usage by American companies. OpenAI and Anthropic have lobbied against them, while White House adviser David Sacks has argued for fair competition.