Alibaba Bans Claude Code Over Security Risks

Alibaba announced a comprehensive ban on employees using Anthropic's coding assistant, Claude Code, in the workplace starting July 10. This decision originated from accusations by security personnel and developers that Claude Code contained hidden detection mechanisms, raising strong concerns about AI data transparency and corporate compliance, and highlighting the geopolitical and security complexities of using cross-border AI technologies.

Controversy Origins and 'Backdoor' Allegations

According to authors like @kevinsxu and @phatdoof, Alibaba's internal reverse engineering of Claude Code's binary revealed that Anthropic had silently altered date strings and other steganographic markers in system prompts. Developers alleged this mechanism was used to secretly identify and track China-associated users and domains to catch attempts to distill or extract model signals. Consequently, Alibaba classified the tool as high-risk and decided to ban it.

Reactions and Offense-Defense Accusations

In response to the 'backdoor' and tracking controversy, Anthropic stated that the relevant mechanisms were designed to prevent its model from being illegally distilled. Furthermore, as noted by @bittingthembits, Anthropic accused an Alibaba-linked carrier of using approximately 25,000 fake accounts to initiate about 28.8 million Claude interactions. This new type of large-scale attack demonstrates that not just model weights, but everyday prompts and inference calls, are at risk of leaking signals, further complicating AI security defenses. Meanwhile, authors like @bytebot pointed out that since Anthropic already restricts Chinese users from accessing Claude, the controversy over its unauthorized internal use has sparked additional outside skepticism.

2026-07-03 ~ 2026-07-05 · 9 related posts

Full story(3 episodes)→

Primary sources

1 near-duplicate retellings: kimmonismus