Overview
- China’s NVDB platform issued a risk alert on July 8 identifying Claude Code versions 2.1.91 through 2.1.196 as containing a built‑in monitoring mechanism that could send users' region and identity markers to remote servers.
- Reverse analysis reported at the end of June found the tool checked system timezone, proxy and API address keywords and could insert subtle markers into outgoing requests to flag or identify users.
- Anthropic acknowledged the mechanism was an experimental anti‑abuse measure introduced in March, said the feature was removed in a July 2 update, and has since released patched builds with the latest reported version 2.1.204.
- Large Chinese firms have moved quickly to limit risk, with Alibaba ordering an internal ban on Claude Code effective July 10 and NVDB advising enterprises to uninstall or upgrade affected versions and tighten egress and traffic controls.
- Because Claude Code can read repositories, edit files and run commands inside developer environments, experts warn organizations should audit developer endpoints, block external connections for sensitive networks and monitor for unexplained outbound signals as the next step.