AI EngineerGuide

Claude Code's hidden markers

by Ashik Nesin Ashik Nesin

Thereallo has written an interesting blog post about how Claude Code adds subtle hidden markers in system prompt if you use API base URL (which matches specific AI labs) or from chinese timezone like Asia/Shanghai or Asia/Urumqi.

It essential becomes a prompt steganography where the data is hidden in plain sight.

If you’re normal Claude Code user then your system prompt will have date like this

Today’s date is 2026-06-30.

However if those conditions are met then they change the following

2026-07-01-at-23.12.27.png

2026-07-01-at-23.09.59.png

From Anthropic’s side they claim it is to prevent distillation attack or the resellers which we’re hearing a lot recently.

CC silently alters the system prompt using invisible-ish Unicode markers. It encodes proxy / gateway classification into a sentence that looks like plain English. It hides the domain list behind XOR and base64. This is not a malicious feature, but it is a weird choice for a developer tool that asks for trust.

Resource

Stay Updated

Get the latest AI engineering insights delivered to your inbox.

No spam. Unsubscribe at any time.