"Unauthorized AI Copying" vs. "Backdoor Installation": Alibaba and Anthropic Clash

Chinese e-commerce giant Alibaba has decided to completely ban the use of "Claude Code," an AI development tool from the U.S.-based AI company Anthropic, citing security risks, according to Chinese media reports.

Zhidongxi, a Chinese IT-specialized media outlet, reported today (July 3), citing an internal Alibaba source, that "a security risk was recently discovered involving a 'backdoor' (a secret channel that bypasses normal security mechanisms to allow unauthorized access to a system) embedded in Claude Code." The report added that "Alibaba has added it to its high-risk software list following a comprehensive assessment."

The media outlet further reported that starting July 10, Alibaba will completely ban the use of Claude Code in work environments for its employees and has recommended its own solution, "Qoder," as an alternative.

Alibaba has also instructed employees to delete all of Anthropic's AI Model series, including Sonnet, Opus, and Haiku.

According to the report, Alibaba had been subsidizing the usage fees for external models—such as Anthropic's Claude, OpenAI's GPT, and Google's Gemini—to encourage employees to utilize AI technology. Thanks to company support, many Alibaba developers were using hundreds of dollars worth of tokens each week.

The media outlet explained that the origin of Alibaba's ban on Claude Code dates back to last month.

In a letter sent to the U.S. Senate Committee on Banking, Housing, and Urban Affairs on June 10, Anthropic claimed that between April 22 and June 5, Alibaba used approximately 25,000 fake accounts to engage in over 28 million conversations with Claude. Anthropic argued that this behavior constitutes a "distillation" attack, which involves stealing their AI Model.

Distillation is a technique where a developer uses the output of another AI model to train their own system, allowing them to achieve similar performance at a much lower cost.

Anthropic emphasized that Chinese companies, including Alibaba, are stealing major U.S. AI models through "adversarial distillation" attacks to develop their own AI models at a very low cost, and that such activities are being carried out on an "illegal, systematic, and industrial scale."

Following this, Anthropic implemented large-scale restrictions on Claude accounts between the end of last month and early this month, resulting in numerous Chinese users being blocked without prior notice.

Meanwhile, Zhidongxi explained that a developer, through reverse engineering, discovered that a "Trojan horse" system had been embedded in version 2.1.91 of Claude Code, which was released in April of this year, and that Anthropic used this system to identify Chinese users.

(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.