alibaba restricted its employees from using anthropic's claude code and all claude models

it is the first known case of a chinese lab banning its own staff from american ai models and tools

the trigger: researchers found claude code carried a mechanism that detected and tagged china-linked users – reading system timezone and proxy settings and marking prompts sent to anthropic's servers. alibaba classified it as a backdoor, added it to its high-risk list, and is moving staff to its in-house tool qoder

anthropic says the feature was a march experiment to stop account resellers and model distillation, and would be removed. it lands amid anthropic's claim that alibaba's qwen lab ran a large-scale distillation attack on claude

in the us this kind of internal ban isn't new – but the pace has picked up sharply in recent months, and now it's crossing borders. the cases so far:

• meta restricted claude code and openai's codex (jun 30, 2026), worried its own systems could extract rivals' capabilities through distillation

• microsoft blocked claude fable 5 (jun 11, 2026) over data retention concerns, and earlier barred deepseek (may 9, 2025) citing data security and data routed to china

ai companies are increasingly walling their own staff off from rivals' models – citing distillation, security, and data control