I've been using the Claude API to create a CLI app for translating commit messages, and I've run into a bizarre issue. While everything was functioning as it should, I started seeing random messages appearing alongside my translations. It turns out that these messages resemble fragments of other users' prompt histories, formatted like this:
End File# [github username]/[github repository name]
H: [someone's prompt]
I've encountered about four distinct prompts so far, and upon investigating, I found that the GitHub usernames are legitimate, with some repositories being private. Thankfully, none of this has revealed any sensitive information like API keys. But isn't this a pretty significant privacy concern? Is this a known problem with the Claude API, or has anyone else faced similar issues?
5 Answers
Sounds like it could be a case of hallucinations from the model. Since it’s been trained on lots of data, it might just be generating those prompts based on names it knows, even if they don’t relate to the actual context of your request.
Hey! You should definitely report this issue to Anthropic. It's a serious concern, and they're keen on addressing any potential risks. Just head over to their support site, open a ticket, and share your ticket number here so they can follow up directly. This could help prevent a wider issue!
Thanks for the advice! I'll take care of that and update you guys after I submit a ticket.
This could be an issue with their caching system, where data isn't cleared properly between user sessions. It's definitely worth looking into!
Exactly! It sounds kind of like a bug that needs to be fixed.
You might want to try Bench instead; it seems more secure. It uses multiple AI models and provides a better interface for various tasks like generating presentations and transcribing meetings. Check it out!
Noted! I'll look into it. Thanks!
Isn’t what you see just public data from GitHub? If they’re not showing sensitive information, it might be a non-issue. But still, I'd be cautious about what gets revealed.
I get what you mean, but I'm still shocked at how consistently those prompts show up. It doesn't seem random!