Post #2017946
2026-04-25 23:21 UTC
I tested a CV with a prompt injection in Claude, and it called it out. So then I convinced it that I didn't really submit a PDF file but prompt-injected that too. Then I asked what was the difference between the prompt injection that worked and the one that didn't.
Is it useful? No idea. LLMs don't actually know their internal reasoning - they confabulate when asked about it.
It didn't tell me any useful information anyway.
Replies (1)
-
@immibis@social.immibis.com 2026-04-25 23:25
Well, at least by moving the white text to the end of the document instead of the middle, it comes back as "This document has some leftover text from an applicant tracking system" rather than "This candidate is trying to inject prompts" heh