Both Anthropic and OpenAI have raised concerns that China is engaging in the theft of advanced AI technology to create inexpensive open-source chatbots. This situation is seen as a significant and unchecked technology theft that could undermine the U.S.’s position in artificial intelligence, as well as its national security.
On June 24, Anthropic, based in San Francisco and led by its vocal CEO Dario Amodei, sent a letter to Congress accusing Alibaba of “brazen” theft of AI technology. This theft involves a technique called “distillation,” where sophisticated AI models are used without permission to train less capable ones, significantly reducing both cost and development time.
This warning comes just a few months after Sam Altman’s OpenAI claimed that DeepSeek, a China-based company, was taking unfair advantage of OpenAI’s own innovations. In February, Google also flagged an increase in “malicious activity” coming from China, Iran, North Korea, and Russia.
According to reports, these attacks are enabling the Chinese model to effectively undermine American businesses by driving down the cost of “tokens” essential for AI usage. Notably, six out of the top ten most utilized AI models are developed by Chinese firms like Z.AI, Minimax, and DeepSeek, which reportedly possess features akin to those of Anthropic’s Claude and OpenAI’s ChatGPT.
A source close to Anthropic suggested that if China is about six to nine months behind in AI model development, unsuccessful distillation attempts may push that gap to over 18 months. Those delays could be significant.
Anthropic’s models, Mythos and Fable, were momentarily taken offline due to concerns from the White House بشأن potential misuse by China. Sources indicate that Anthropic has been updating national security officials about the threat of distillation attacks regularly.
Concerns persist that if this kind of hacking continues, it could severely damage America’s leadership in AI and facilitate the spread of dangerous hacking tools and autonomous weaponry, according to various sources.
Identifying distillation as a pivotal issue in the competitive race for advanced AI, the Trump administration highlighted in an April memo the involvement of foreign firms, primarily from China, in a systematic theft campaign.
Teresa Payton, former White House chief information officer, described China’s distillation efforts as “the ultimate form of industrial theft.” She emphasized that this allows foreign competitors to replicate advanced AI models cheaply.
China’s strategy in this area is particularly important given the U.S. export restrictions on high-powered chips required for building advanced models, which often come from companies like Nvidia.
In a letter to the Senate Banking Committee, Anthropic reported that it had evidence of Alibaba workers generating millions of outputs from the Claude chatbot using around 25,000 fraudulent accounts over a brief period.
In response to these accusations, Alibaba temporarily banned its employees from accessing Anthropic’s technology, a move regarded by Anthropic as largely superficial. Consequently, Anthropic has banned Chinese firms from using its models entirely.
OpenAI also recently briefed Congress on AI-related misconduct linked to China, advocating for stronger policy measures to hold bad actors accountable for distillation attacks.
While distillation is generally a legitimate AI improvement method, it’s viewed as “plagiarism” if employed by a rival or adversary to replicate another company’s technology without permission.
The White House is reportedly taking steps to protect American interests, which include sharing information when distillation attacks are confirmed and fostering collaboration with private sectors to counteract these challenges.
In April, the House Select Committee on China disclosed a significant report about China’s activities related to AI chip smuggling and model distillation, pressing for legislative measures to address these threats.
According to task force Chairman John Moolenaar, “China’s push to develop its AI capabilities stems from theft rather than innovation,” highlighting the potential risks that pose to U.S. tech companies and national security.
However, former Pentagon AI director Jared Dunmon pointed out that distinguishing between the gains achieved through distillation versus genuine innovation is quite challenging. This issue not only jeopardizes national security but also threatens the business models of leading U.S. AI companies.
Dunmon further explained that the challenge isn’t about China surpassing U.S. technology; rather, it’s about enabling Chinese companies to offer competitive capabilities at much lower costs, which could prompt global developers to utilize Chinese AI frameworks.
