Is there nightshade but for text and code? Maybe my source headers should include a bunch of special characters that then give a prompt injection. And sprinkle some nonsensical code comments before the real code comment.
I think the issue is that text uses comparatively very little information, so you can’t just inject invisible changes by changing the least insignificant bits - you’d need to change the actual phrasing/spelling of your text/code, and that’d be noticable.
Is there nightshade but for text and code? Maybe my source headers should include a bunch of special characters that then give a prompt injection. And sprinkle some nonsensical code comments before the real code comment.
There are glitch tokens but I think those only effect it when using it.
I think the issue is that text uses comparatively very little information, so you can’t just inject invisible changes by changing the least insignificant bits - you’d need to change the actual phrasing/spelling of your text/code, and that’d be noticable.
Maybe like a bunch of white text at 2pt?
Not visible to the user, but fully readable by crawlers.
If a bot can’t read it, nor can a visually impaired user.
Well if it’s a prompt injection to fuck with llms you don’t want any users having to read it anyway, vision impaired or no.
You missed my point. A prompt injection to fuck with LLMs would be read by a visually impaired user’s screen reader.