Alibaba is banning Claude Code over backdoor claims: Here’s why

Alibaba will bar employees from using Anthropic's Claude Code inside company workspaces starting July 10, according to a source familiar with the matter cited by Reuters, over allegations that the coding tool secretly embedded detection logic targeting Chinese users.

The accusation can be traced back to a Reddit post on Tuesday, made by LegitMichel777, claiming that they discovered some code obfuscation during the restoration process of the remote-control function in Claude Code.

According to the report, since the 2nd of April, version 2.1.91 of the software was checking whether the user’s computer had Asia/Shanghai or Asia/Urumqi timezone settings or if there was an open proxy with a URL matching one of the 147 entries of a decoded list of Chinese technology sites, cloud regions, and AI labs like Alibaba, Baidu, ByteDance, and Moonshot AI.

Instead of signalling telemetry through a visible flag, the alleged mechanism modified some minor machine-readable parameters in the system prompt of Claude Code, changed the date format from dash-separated to slash-separated and replaced the apostrophe in "Today's date is" with the visually similar Unicode character.

Thariq, an employee of the Claude Code Team at Anthropic, clarified through social media that this mechanism was meant to stop the selling of accounts and the distillation of models and not spying on users, adding that the mechanism would be taken down in the next update. The Register and several other media outlets noted that this mechanism was being rolled back as of July 1; it was in operation for about three months.

The ban comes as tensions were already rising between the two companies. In a June 10 letter sent to US senators, Anthropic accused operators from Alibaba's Qwen AI lab of using close to 25,000 fraudulent accounts to syphon out the programming and reasoning capabilities of Claude through more than 28.8 million conversations from April 22 to June 5.