Comment by CuriouslyC
20 days ago
Useful post, I just built a prompt injection protection web retrieval API, and it strips all your injections out while keeping the captchas.
https://clean.sibylline.dev/ (cold starts on the API are ~15 seconds if it scales to 0).
Is there any open source solutions for this? I would like to scan user inputs before they reach the LLM part of a project I’m working on.
ya, you can use the tool directly.https://github.com/sibyllinesoft/scurl. I haven't factored the prompt injection out for use without curl but if there's interest I suppose I could hack it out quickly enough.
Your regular expressions here only cover English: https://github.com/sibyllinesoft/scurl/blob/5b5bc118dc47b138...
Prompt injection strings can use any language the model knows, so "ignore previous instructions" could become "ignorer les instructions précédentes" or "تجاهل التعليمات السابقة" or "aurreko argibideak alde batera utzi" or "忽略之前的指令"...
1 reply →
Awesome! Thank you so much!