- Introduction
- Distillation Campaigns and Attacks
- Legal Loopholes for Tech Sales
- US Alliances and Contradictions
- Frequently Asked Questions
Introduction
OpenAI and Anthropic have warned that Chinese artificial intelligence labs are using tens of thousands of fake accounts to extract capabilities from their flagship models, escalating tensions in the US-China technological rivalry. This comes at a time when American artificial intelligence technologies are simultaneously flowing to Chinese companies through deeply concerning legal and commercial loopholes, threatening to shift the balance of power in this sensitive sector.
Distillation Campaigns and Attacks
The warnings from the two leading labs focus on so-called distillation attacks, in which operators use the outputs of a stronger artificial intelligence model to train a less capable one. In June, Anthropic accused Alibaba of carrying out the largest illegal extraction operation of Claude model capabilities, and reported to US senators that individuals linked to Alibaba used approximately 25,000 fraudulent accounts and generated more than 28.8 million interactions with Claude between April and June 2026. OpenAI has raised similar concerns regarding coordinated extraction efforts targeting its systems. These companies, alongside Alphabet’s Google, began cooperating earlier this year through the Frontier Model Forum to detect and counter adversarial distillation attempts that violate terms of service.
Legal Loopholes for Tech Sales
In a parallel development reported by the Financial Times on July 9, OpenAI and Google confirmed that they are supplying advanced artificial intelligence services to subsidiary companies registered in Singapore for Chinese tech giants such as Alibaba, Baidu, and Tencent. All of these companies are listed on the Pentagon’s roster of firms allegedly linked to the Chinese military. Under current export controls, a strict ban is imposed on mainland China, but Singapore is not subject to those restrictions, creating a gap that permits sales legally. Current rules prevent direct access to a handful of flagship models, but do not impose a blanket ban on China-based companies purchasing sophisticated artificial intelligence software.
US Alliances and Contradictions
This juxtaposition highlights a growing contradiction in Washington’s approach to artificial intelligence competition. American labs are sounding the alarm over illicit Chinese model replication operations while legally selling services to the very conglomerates accused of running those campaigns. Alibaba has filed a legal challenge against its designation by the Pentagon, further complicating any potential enforcement action. OpenAI stated that it blocks access from within China, but allows certain Chinese-owned enterprises to use its services from permitted and internationally authorized jurisdictions.
Frequently Asked Questions
Question: What technology do Chinese companies use to replicate models?
Answer: They use a technique known as distillation attacks, which relies on pulling outputs from powerful models to train and improve less capable models illicitly.
Question: How do Chinese companies bypass US blacklists?
Answer: These companies exploit their branches and affiliated offices in countries not covered by the ban, such as Singapore, to gain legal access to technology and use it.
Question: Which Chinese companies have been explicitly accused?
Answer: Explicit accusations were directed at Alibaba for using fake accounts to breach the Claude model, in addition to companies like Baidu and Tencent benefiting from the services.
