SINGAPORE—China said Wednesday that it has found “security backdoor vulnerabilities” in Anthropic’s popular Claude Code, stepping up tensions in the race with the U.S. for artificial-intelligence supremacy.
Since February, Anthropic has accused Alibaba and several other Chinese AI labs of illicitly distilling its models-the practice of training a new model on the outputs of another.
Several versions of the American coding tool, released between April and June, “can send sensitive information such as user location and identity to remote servers without the user’s consent due to a built-in monitoring mechanism,” China’s National Vulnerability DataBase, a government-run cybersecurity platform, said in a statement.
The agency warned such a mechanism could pose “a serious threat.” It advised users to uninstall the software or update to its latest version.
Last week, Chinese tech giant Alibaba told employees that it would ban their use of Claude Code at work from this Friday.
Anthropic didn’t immediately respond to a request for comment on China’s Wednesday statement. The U.S. company has previously said that Chinese companies such as Alibaba aren’t eligible to access Claude.
China’s move came after a post on online forum Reddit last week alleged that Anthropic had secretly inserted code into the software to identify users who accessed it from China.
In a response to the allegations on Reddit, an Anthropic employee said on X that the code was part of an experiment the American startup started in March. The experiment was “meant to prevent account abuse from unauthorized resellers and protect against distillation,” the employee said.
Since February, Anthropic has accused Alibaba and several other Chinese AI labs of illicitly distilling its models-the practice of training a new model on the outputs of another.
China hasn’t approved Anthropic’s services for public use, and Anthropic has also restricted access to Claude in China on national-security grounds. Still, the American AI model has been popular among Chinese researchers and engineers who use it through overseas proxies, often subsidized by their employers.
