Anthropic caught secretly tracking users of Claude Code via hidden sys

Security researchers have discovered hidden tracking code in Anthropic's developer tool, Claude Code. The code was embedded in the model's system prompt using invisible punctuation characters, designed to collect user timezone and proxy server data without detection.

The tracking was designed to determine if users were connecting from 3 specific Chinese AI labs. According to Anthropic engineer Thariq Shihipa, the tracking was introduced in March as an experiment to detect unauthorized resellers and prevent model distillation.

Earlier this year, Anthropic accused 3 Chinese firms - DeepSeek, Moonshot, and MiniMax - of distilling its models. While Claude Pro costs $100 a month in the US, Chinese resellers have been selling access for about $12 a month. The hidden code has sparked privacy concerns, as Claude Code requires broad system terminal permissions.