Lobsters0 Hot0 bình luận5 phút đọc3 giờ trước

Claude Code Is Steganographically Marking Requests

A developer inspecting Claude Code (v2.1.196) for privacy reasons discovered that the tool silently alters its system prompt based on the ANTHROPIC_BASE_URL environment variable and system timezone. Specifically, the apostrophe in "Today's date" is replaced with different Unicode characters (U+2019, U+02BC, or U+02B9) depending on whether the API hostname matches a decoded domain list or contains AI lab keywords, and the date separator changes from '-' to '/' for Asia/Shanghai or Asia/Urumqi timezones. The domain and keyword lists are obfuscated with base64 and XOR encoding. The author argues this is prompt steganography — encoding classification signals invisibly into plain-looking text — likely intended to detect API resellers or model distillation pipelines. While not malicious, the hidden behavior undermines trust in a tool that already has broad filesystem and shell access, and the author argues Anthropic should make such telemetry explicit and documented rather than hidden.

Đọc bài gốc

#privacy #claude-code #reverse-engineering #prompt-injection

Nguồn: https://thereallo.dev/blog/claude-code-prompt-steganography. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.