Note by @fastfinge

Tamas G @Tamasg@mindly.social

Aha. So Claude's going to start leaking random remote access URLs to people's Claude Code sessions now, I see we're playing that game. Nope nope. Not sure what prompted this new behavior all of a sudden, but every single release or document it writes gets stamped with the "generated with Claude Code" ending, and that slug link points to directly someone's session ID they used for the coding session itself. Just wow.

Tamas G @Tamasg@mindly.social

I mean, If a government entity subpoenas Anthropic for information on a specific session ID found in a repository, Anthropic can instantly pull up the exact prompts, system context, and chat history that created that code. I don't ccare that people can't take control or see the history of the session and there's security. You are taking a secret token (the session ID) and treating it like public metadata. Hackers love public identifiers. An attacker doesn't need to break into your specific computer today. Instead, they can write a simple script to scrape GitHub for any link matching claude.ai/code/session_*. They can build a massive database of active or historical session IDs. If an attacker finds a zero-day exploit (an unpatched vulnerability) in Anthropic’s session routing, web authentication, or cross-origin headers in the future, they don't have to guess random session tokens. Anthropic already gave them that data as a free gift.

🇨🇦Samuel Proulx🇨🇦 @fastfinge@interfree.ca

@Tamasg Right, but if you're putting the code out in public, the contents of your session should be public safe anyway, in theory. In reality, anything you give to any AI anywhere is something you should be fine with the public consuming either due to a zero day or as part of training data anyway.

Tamas G @Tamasg@mindly.social

@fastfinge what makes me queezy on that one though is that it still contains private paths to files and codebase, programs, or potentially token storage in user folders - the actual names of certificates, so while that can't entirely leak sensitive data itself that decrypts the certs or has the direct token in context, it can still give someone an incredible layout-view of a subset of someone's folders, some of which may live outside-repo. So it's just a very, very slippery sloap. (and why I tell Claude Code to always scan the repo for any stray local paths that may leak.)

🇨🇦Samuel Proulx🇨🇦 @fastfinge@interfree.ca

@Tamasg But all that stuff is already being fed to the cloud, and into the training data. Once Claude has touched it, it's not private, anymore.

Tamas G @Tamasg@mindly.social

@fastfinge nah, you can disable data training though, which a conscious choice I made - and I always get asked if Anthropic can look at my transcripts or not. Most times I say no, unless it's something I know won't have sensitive info like dirs from old CD-ROMS and such.

🇨🇦Samuel Proulx🇨🇦 @fastfinge@interfree.ca

@Tamasg If you believe any of the major AI companies are respecting that setting, I have a bridge to sell you.