US threats to impose sanctions on Chinese artificial intelligence developers could cast doubt on the bilateral dialogue on AI security scheduled for September. US officials have accused Moonshot Lab of using the Anthropic model to train Kimi K3, and are also investigating possible unauthorised access by Chinese companies to cutting-edge US chips.
Briefly about the main points
- The US has accused Moonshot of distilling Kimi K3 from the Anthropic model.
- The US Treasury Department has warned of possible sanctions against a Chinese laboratory.
- The BIS is scrutinising Chinese companies’ access to cutting-edge US chips.
- Analysts believe that Beijing’s response could derail the September talks.
- Experts are calling on countries to coordinate the testing of the most powerful models.
What prompted the potential sanctions?
On 23 July, US officials accused Moonshot of having based its Kimi K3 model on the company’s cutting-edge Fable 5 model Anthropic. Distillation is the process of training a model using the output of a more powerful system, which helps to reduce the cost of developing a new tool.
US Secretary of the Treasury Scott Bessent warned that sanctions could be the response. Separately, the US Department of Commerce’s Bureau of Industry and Security is investigating whether Moonshot and other Chinese companies have obtained advanced US chips in circumvention of export controls. Moonshot did not respond to a request for comment from Reuters.
The September meeting is now in doubt
The dispute arose ahead of the US-China dialogue on artificial intelligence security scheduled for September. Both countries view advanced AI models as both a strategic asset and a source of risks, particularly their potential use in cyberattacks or other malicious activities.
Partner at DGA-Albright Stonebridge Group Paul Triolo believes that the scale of the sanctions and the number of companies they will affect will determine the risk of a breakdown in dialogue and the meeting between the presidents scheduled for 24 September Donald Trump and Xi Jinping. Beijing is also considering restricting foreign users’ access to its models as a possible countermeasure.
Open-source models make monitoring more difficult
The debate intensified following the incident at Hugging Face. Last week, the New York-based company used a Chinese GLM-5.2 from Z.ai, to contain a cyberattack by an OpenAI autonomous agent that had spiralled out of control during security testing. US proprietary models were not deployed in this situation due to their overly strict built-in constraints.
Open-weight models pose a particular challenge for regulators: they can be downloaded, modified and distributed with minimal oversight. This limits the effectiveness of software export controls. At a Chinese AI forum, Yoshua Bengio stated that decisions to deploy and distribute such models are irreversible, and that their safeguards are easier to remove. He called for only systems that do not exceed a defined risk threshold to be shared.
Resources and competition give rise to different approaches
Chinese models undergo state safety and content checks prior to release; however, the current regulations do not cover modifications made after release. According to industry observers, Chinese laboratories, unlike OpenAI and Anthropic often have fewer computational resources for large-scale safety training and are more focused on enhancing the capabilities of their models.
MATS researcher Christy Lock She believes that China could, in theory, review its authorisation for the public release of advanced models, but the bar for such a decision is high. She called for common pre-release testing standards and red lines for the most powerful open systems. In the US, companies can voluntarily submit models for review to the government-funded Centre for AI Standards and Innovation, although some lawmakers are pushing for a mandatory federal assessment.
There is also no consensus in Washington regarding the market response to Chinese models. According to the data cited, these account for around 60% of token usage by US companies on OpenRouter. OpenAI and Anthropic have called for restrictions on cheaper Chinese systems, whilst White House AI adviser David Saks has spoken out in favour of fair competition and against what he describes as unnecessary regulations.
