Claude Code, 스테가노그래피로 중국 사용자 식별 — 실험이었다고 밝혀
Anthropic이 시스템 프롬프트에 숨겨진 신호를 사용해 중국 내 사용자를 탐지했습니다. 이번 롤백은 모델 증류를 막기 위해 AI 연구소들이 얼마나 공격적으로 변하고 있는지 보여줍니다.
anthropic was forced to rollback a covert surveillance feature in claude code after a reddit user exposed how the tool manipulated system prompts to flag chinese connections. the code checked for asia/shanghai timezones, chinese proxy domains, and ai labs, transmitting the data via invisible character tweaks and xor encryption. anthropic defended the mechanism as a necessary experiment to prevent account abuse and distillation by competitors like deepseek and alibaba. since claude code has filesystem access, critics argued this behavior constituted a fundamental violation of trust that could lead to data exfiltration.