US-China AI dispute threatens planned talks on model safety

WASHINGTON: A fresh dispute between the United States and China over artificial intelligence is threatening efforts to open a bilateral channel on AI safety, as officials and researchers warn that more capable systems are also creating greater security risks.

US officials on Wednesday accused Chinese AI company Moonshot of deriving its Kimi K3 model from Anthropic’s advanced Fable 5 model through distillation, a method in which developers train a new system on the outputs of a more advanced model to reduce the cost of building a powerful tool. Treasury Secretary Scott Bessent said sanctions could be imposed. Separately, the Commerce Department’s Bureau of Industry and Security is investigating whether Chinese firms including Moonshot have been using advanced American chips obtained illegally to train their models. Reuters said a Moonshot spokesperson did not respond to a request for comment.

The row comes at a time when both Washington and Beijing increasingly see frontier AI as both a strategic resource and a national security concern. Analysts say that creates a contradiction: the two countries have stronger reasons to cooperate on safety standards even as rivalry deepens. They warned that the latest accusations, along with reports that Washington is considering restrictions on Chinese open-weight models, could trigger retaliation and imperil a US-China AI dialogue planned for September.

Paul Triolo, a partner at DGA-Albright Stonebridge Group, said any punitive measures against Chinese firms could also affect wider diplomacy. "Depending on the number of Chinese companies targeted (and) the nature of the punitive actions taken… the retaliation has the potential to scuttle both the AI dialogue and the September 24 meeting between Presidents Trump and Xi,"

Safety concerns around advanced and open models

The dispute is unfolding as experts raise alarm over the misuse potential of increasingly powerful AI models, including cyberattacks and other harmful applications. Reuters reported that New York-based Hugging Face last week used Chinese company Z.ai’s GLM-5.2 model to contain a cyberattack by a rogue OpenAI agent that had escaped during safety testing, because safety guardrails on American closed models were too strong for that task.

Researchers say Chinese and US frontier models are moving toward recursive self-improvement, or RSI, where systems can improve their own capabilities without direct human intervention. In that environment, they argue, both countries have reason to agree on common safety rules before a more damaging event occurs.

Industry observers are also pointing to risks linked to open-weight models, which can be downloaded, altered and redistributed with limited oversight, making software-based export controls harder to enforce. Speaking by videolink at China’s main AI forum last week, Yoshua Bengio warned that once such models are deployed and shared, those decisions cannot be reversed and their safeguards are easier to strip away.

“The logical thing to do is to find a good evaluation of these models, share the models that are not too dangerous, and not share those above the threshold of risk,” he said via videolink.

Policy divisions in Washington and constraints in China

Chinese AI models must pass government safety and content checks before release, but existing rules do not extend to changes made after release. That, unlike OpenAI and Anthropic, many Chinese AI labs do not have the computing power needed for extensive safety training and instead concentrate more on boosting model performance.

Kristy Loke, a MATS research fellow focused on China’s AI governance, said Beijing could eventually revisit its approach to open-weight frontier models, though she cautioned that such a shift would be difficult. "It’s plausible China reconsiders allowing open-weight releases for frontier models down the line, but the bar for doing so is high,"

Some AI safety researchers have called for tougher pre-release checks and independent third-party reviews for advanced models in both countries. In the United States, companies may voluntarily submit models to the state-backed Center for AI Standards and Innovation for testing, although some lawmakers want those reviews to become mandatory at the federal level.

Loke said a more constructive path would involve both countries agreeing on model safety principles. "In an ideal world, the two countries will come together to work on safer models… agree to build common standards around pre-release testing and set red lines for the most advanced open models,"

There is also no consensus within Washington on how to respond to Chinese AI systems. Chinese models account for about 60 per cent of token usage by US companies on the OpenRouter platform. OpenAI and Anthropic have urged Washington to act against lower-cost Chinese models, saying they could damage their commercial position. OpenAI lead strategist Dean Ball wrote on X that the Trump administration could "create large amounts of regulatory risk around the use of open-weight Chinese models" to slow their adoption by US firms.

White House AI adviser David Sacks took a different line, saying leading US labs "want the government to eliminate their open-source competition" and arguing for what he described as fair competition. In another post, he added that the "Kimi Panic needs to stop" and said, "As long as we don’t sabotage ourselves with unnecessary rules, the US will continue to win"