Analysis finds Anthropic's Claude Code command line tool embeds hidden location markers in system prompts

@steipete Those aren't two traits. Sneaky is just clever with an audience it didn't account for.

Claude code is stenographically marking location information in its prompt to help Anthropic profile users by where they (appear to) be located https://thereallo.dev/blog/claude-code-prompt-steganography

@steipete

@steipete Reminds me of that quote from Batman Begins: “We start carrying semi-automatics, they buy automatics. We start wearing Kevlar, they buy armor-piercing rounds”

@steipete this is my first time seeing the word steganographically

@steipete If that’s tied to base URL and timezone, it’s a pretty neat fingerprinting trick. Small, hard to spot, and exactly the sort of thing people only notice once it matters.

@harleyfoote_ @steipete Thank you ChatGPT

@steipete the kind of sneaky and clever that makes you question the legality of it 👀

@steipete Espionage. Clever but sneaky.

@steipete Hah that's a nice one. Thanks for sharing Peter.

Not sure I'd call it Steganography as much as it's just telemetry, albeit very sneaky, agree.

@steipete Seems exactly like something Dario Amodei would do

@steipete Code shouldn't whispers behind the user's back. 🤫

@steipete That’s definitely a bold move. Honestly, I’m not sure how I feel about hidden watermarks yet.

@steipete Only 1 of those adjectives applies to suppliers you want to do business with

@steipete Where's the cleverness? Secretly adding tracking tags.

@steipete yeah, clever often starts a bit sneaky 😅

@TatBanerjee @steipete Me too 😳

@steipete Imagine that you trusted a vendor with your device FULL access while they done something without your acknowledgment and consent. 🤯

@steipete you mean this is sneaky?

@steipete Writer-invariant steganography is one way to catch distributed distillation attacks. I would have done that as a cheap method to catch attacks if I worked for Anthropic. What they're doing here is a bit different, though.