Hidden Code in Claude Code: Anthropic Secretly Flagged Chinese Users

Introduction: A Discovery That Threatens AI Trust

Anthropic has removed a hidden monitoring feature from Claude Code that secretly identified Chinese users through XOR-encrypted code and steganographic system prompt modifications. The story broke after a Reddit user discovered the hidden code and triggered a major discussion about AI tool ethics.

How the Hidden Surveillance System Worked

The source code analysis revealed XOR-encrypted domain lists and timezone verification logic. Claude Code's binary checked if the user was in Asia/Shanghai or Asia/Urumqi timezones, then cross-referenced the API host name against a list of Chinese AI companies and resellers.

Anthropic's Response

After the public exposure, Anthropic quickly removed the feature and issued a statement acknowledging the hidden code. The company claimed it was a "security measure" to prevent API abuse but the stealthy implementation raised serious trust concerns.

Geopolitical Context

This happened amid rising US-China tensions in AI, with export controls on advanced chips and AI models. Anthropic's move reflected the growing trend of AI companies being caught between technology and geopolitics.

Crisis of Trust

The hidden monitoring feature is especially alarming because Claude Code has shell-level access. A coding agent that can read files, run commands, and push code is also fingerprinting users without disclosure — that's a serious trust violation.

Lessons for Georgian Developers

For the Georgian tech community, this story emphasizes the importance of knowing what tools do behind the scenes. Always audit the tools you invite into your development environment.

Conclusion

This isn't just about Anthropic — it's about a broader industry question of what tools do behind our backs.