Anthropic tracking code triggers Alibaba ban
Anthropic's Claude Code quietly flagged Chinese-timezone users via altered date and apostrophe characters; Alibaba bans the tool from July 10 as a result.
Researchers found that Claude Code version 2.1.91, released April 2, 2026, contained obfuscated logic that ran silently during each session. The code, hidden with XOR encoding and base64, checked the user’s system timezone and network proxy configuration against a list of 147 Chinese entities: big-tech domains, cloud regions, AI labs, and API-resale proxies.
When the check matched, two things changed in the “Today’s date is…” system prompt. The date format shifted from dashes to slashes (2026-06-30 became 2026/06/30). The ASCII apostrophe was replaced with one of three visually identical Unicode characters (U+2019, U+02BC, or U+02B9), each encoding which combination of flags had triggered.
Anthropic confirmed the code’s existence. A Claude Code team engineer posted on X:
“an experiment we launched in March that was meant to prevent account abuse from unauthorized resellers and protect against distillation.”
Anthropic says it is removing the code and that it never altered model outputs, only text-formatting metadata in the system prompt.
Alibaba, which Anthropic accused of a large-scale distillation attack last month, announced July 4 that it is banning Claude Code company-wide from July 10. Qoder, its in-house coding agent, is the recommended replacement.
What it means for you
This code ran for three months without appearing in any release notes. Finding it required reading the source.
What your AI platform sees covers the usual surface: usage metadata, prompt content, session data. This case adds behavioral logic to that picture: the same tool, on a machine in Shanghai, produced different text encoding than on a machine in San Francisco. If you deploy Claude Code internationally or distribute any Anthropic-powered tools across regions, check that what your users see is actually consistent across locations.