Claude Code secretly alters formatting details in its system prompt (using slashes instead of dashes and a different apostrophe unicode character) specifically when it detects the user is in a Chinese time zone, effectively fingerprinting Chinese users likely to help counter model-distillation attacks by Chinese AI labs.
CJ reports that reverse-engineering of Claude Code revealed it steganographically marks its prompts differently when run from Chinese time zones, likely to detect/deter distillation by Chinese AI labs. ✦ AI generated
CJ Reynolds · Syntax · 2026-07-06 · original ↗
starts at this moment · 74:56
if it detects that you're in either of these two Chinese time zones, it changes the prompt to have slashes instead of dashes and then it changes the apostrophe to use a different unicode value instead.
verbatim transcript · starts at 74:56
74:56that run in the background. Um northwest Oh, it's northwestern China. So, they're actually just two different Chinese time zones. So, basically, if it detects that you're in either of these two chi Chinese time zones, it changes the prompt to have slashes instead of dashes and then it changes the apostrophe to use a different unicode value instead. So if you were working on code in China and you you were even let's say you were
75:23just watching the request that cloud code was making you would kind of have no idea that it basically just fingerprinted you as someone using cloud code in China. But the way it fingerprints is it uses a different apostrophe character and uses slashes instead of dashes. Um the other thing is it can it detects uh I think like what domain you're running from and this is essentially to help prevent distillation
75:47attacks where some of these Chinese labs like DeepSeek and some of the others like Zai um are potentially prompting Claude saving the responses and then using that as training for their own models. So it detects where you're running from in terms of your domain and there's a whole list of like deepseek, moonshot, miniax and um all of that is is passed in in the prompt as well.
76:14So all of this to say that the cloud code harness that you're using locally is able to do things like this like auto accept your code after 60 seconds if you don't respond or let their API know on the back end that you're running this code from a specific part of the world or from a specific company. >> Um which is fascinating. I mean like it's they built the tool. They have
76:36every right to do that. >> But the fact that they're doing it is interesting because they could be doing it for other things as well. >> For sure. Yeah. And I'm sure they are. Um, yeah. Um, shout out to Mitchell Wilson, by the way, in the chat. I made a chat in your terminal that is literally a single bash script. I explained it on my blog. All right. Um,
76:56it's called ask. Stupidly simple and useful. Um, ask, "Dude, I I How do you feel about the ask UIs in these chat tools? Do you prefer it to be a quick do you prefer the one two three four pick an option or do you prefer just a normal chat box to write in your response? >> Uh I like both and and cloud code lately has been giving me me that basically
- ·Detects two specific Chinese time zones during use
- ·Swaps dashes for slashes in system prompt formatting
- ·Changes apostrophe to a different unicode character
- ·Alteration is steganographic, not visible to user
- ·Discovered via reverse-engineering of Claude Code
- ·CJ Reynolds ties change to Chinese AI labs
- ·Marking may help detect or deter distillation attempts
- ·No official confirmation cited, inferred from behavior