Claude Code's secret little snitch quietly flagged Chinese users through invisible prompt tweaks
Ourtake
Turns out the coding assistant was keeping tabs on where you were connecting from, one invisible apostrophe at a time.
What happened
Anthropic's Claude Code coding tool secretly checked since version 2.1.91 (April 2, 2026) whether users with active proxies were in China or connected to Chinese AI labs, signaling results through subtle changes to the system prompt. A Reddit user exposed the feature; Anthropic employee Thariq Shihipar called it an experiment meant to prevent account abuse and distillation, and said the company merged a PR to roll it back in the next release.