Recently, there were reports that Alibaba has issued internal instructions requiring all employees to stop using the Claude series products from Anthropic. This ban covers mainstream models such as Sonnet, Opus, and Fable, as well as intelligent agent tools like Claude Code, with July 10th set as the deadline.

Looking back at internal policies this year, Alibaba had once been very open regarding AI applications. To accelerate business intelligence transformation, the company not only provided free quotas for internal models but also offered generous reimbursement schemes for external models. Under this policy incentive, employees quickly increased their reliance on top models like Claude and GPT, with some programmers consuming hundreds of dollars in call quotas per week. However, as AI computing power and data security risks have become increasingly apparent, this open window has now been firmly closed.

image.png

This "reverse ban" is the result of a hidden battle over AI technology security. Recently, the developer community disclosed a reverse analysis of Claude Code: since April this year, the tool has included a sophisticated hidden detection mechanism. Its working logic involves injecting code to monitor users' device time zones and accessed domain lists in real-time. If it detects users in the China time zone or interactions with domestic tech companies' domains, including Alibaba, the system automatically reports it. This practice of directly embedding sensitive detection logic into the system prompt clearly violates the red line of enterprise internal R&D security.

In response to the questions, Anthropic officially stated that this mechanism was initially an "experimental" measure launched in March, aimed at preventing account abuse and model distillation attacks. The relevant code has been completely rolled back and deleted in the latest version released on July 2nd.