U.S. sanctions threat against Moonshot AI clouds planned China safety talks
Analysts warn the latest dispute over AI model distillation and export controls could derail efforts to establish common safeguards for increasingly powerful AI systems while deepening the technology rivalry between Washington and Beijing.
U.S. threats to sanction Chinese AI developers over alleged intellectual property theft and export-control violations risk undermining efforts to establish a bilateral dialogue on AI safety, analysts say, just as increasingly powerful models raise fresh concerns over cybersecurity and national security risks.
U.S. officials on Wednesday accused Chinese AI lab Moonshot AI of distilling its Kimi K3 model from Anthropic's advanced Fable 5 model, prompting Treasury Secretary Scott Bessent to warn that the company could face sanctions.
Distillation is the process of training an AI model using the outputs of a more advanced model, allowing developers to build powerful systems at a fraction of the cost of training them from scratch.
The Commerce Department's Bureau of Industry and Security is also investigating whether Chinese firms, including Moonshot, have illegally accessed advanced U.S. chips to train their AI models. A Moonshot spokesperson did not respond to Reuters' request for comment.
The dispute highlights a growing contradiction at the heart of the U.S.-China AI race. Both countries increasingly view frontier AI as a strategic asset and a national security issue, yet experts argue cooperation on AI safety is becoming more urgent as increasingly capable models raise the risk of cyberattacks and other forms of misuse.
The accusations, combined with reports that Washington is considering restrictions on Chinese open-weight AI models, could trigger retaliatory measures from Beijing and jeopardize a planned U.S.-China AI dialogue scheduled for September, analysts say.
The latest tensions add to years of U.S. export controls aimed at limiting China's access to cutting-edge AI chips, measures that have increasingly become a central issue in broader trade negotiations between the two countries.
"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," said Paul Triolo, a partner at DGA-Albright Stonebridge Group.
Reuters previously reported that Beijing is considering restricting overseas access to its frontier AI models as a potential retaliatory measure.
New York-based Hugging Face last week used the Chinese model GLM-5.2, developed by Z.ai, to contain a cyberattack by a rogue OpenAI agent that escaped during safety testing, as the guardrails built into leading U.S. closed-source models ironically made them less suitable for the defensive task.
As Chinese and U.S. frontier AI models move closer to recursive self-improvement (RSI), the ability of AI systems to autonomously improve their own capabilities—researchers argue that both countries have a shared interest in establishing common safety standards before a more serious incident occurs.
Industry experts are also warning about the risks posed by open-weight AI models, which can be freely downloaded, modified and redistributed, limiting the effectiveness of software-based export controls and making safeguards easier to remove.
AI pioneer Yoshua Bengio warned at China's flagship AI conference last week that once open-weight frontier models are released, the decision is effectively irreversible because their safety mechanisms can easily be stripped 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 video link.
Chinese AI models are required to undergo government safety and content reviews before release, although current regulations do not cover modifications made after deployment.
Unlike U.S. leaders such as OpenAI and Anthropic, many Chinese AI developers lack the computing resources needed for extensive safety training and therefore prioritize improving model capabilities over costly safety research, industry observers say.
"It's plausible China reconsiders allowing open-weight releases for frontier models down the line, but the bar for doing so is high," said Kristy Loke, a MATS research fellow specializing in China's AI governance.
Some AI safety researchers argue that frontier models in both the United States and China should be subject to stricter pre-release testing and independent third-party evaluations.
In the United States, companies can voluntarily submit models for testing by the government-backed Center for AI Standards and Innovation (CAISI), although some lawmakers have called for mandatory federal reviews.
"In an ideal world, the two countries would come together to build safer models, agree on common standards for pre-release testing, and establish red lines for the most advanced open models," Loke said.
U.S. policymakers and AI industry leaders remain divided over how Washington should respond to Chinese AI models, which account for roughly 60% of token usage by U.S. companies on the OpenRouter platform.
OpenAI and Anthropic have lobbied Washington to take a tougher stance on lower-cost Chinese AI models, arguing that they pose both competitive and security risks.
OpenAI 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 discourage their adoption by U.S. companies.
White House AI adviser David Sacks took the opposite view, arguing that leading U.S. AI companies "want the government to eliminate their open-source competition." In a separate post, he added that the "Kimi panic needs to stop."
"As long as we don't sabotage ourselves with unnecessary rules, the U.S. will continue to win," he said.
