💻 Claude Code uses steganography for covert query marking
Researchers have discovered a mechanism in Claude Code (v2.1.196) for covertly marking queries through steganography within the system prompt. The tool imperceptibly alters Unicode apostrophe characters and date separator formats to transmit metadata regarding traffic routing and the domains being used.
🌍 This reveals hidden user monitoring methods employed by AI tool developers and raises questions about tool transparency.
👤 If you are using Claude Code via third-party APIs or proxies, your traffic may be classified by Anthropic without your knowledge.
Source 1: https://thereallo.dev/blog/claude-code-prompt-steganography