logoalt Hacker News

verdvermyesterday at 7:54 PM2 repliesview on HN

I don't understand why they don't look for large substring matches for the system prompt before returning the response. Trivial calculation compared to a system prompt instruction asking the model not to do it


Replies

anvuongyesterday at 9:45 PM

because you can always make up your own language and ask the model to use it, no filters would catch that

LoganDarkyesterday at 8:05 PM

Because it's trivial to bypass through things like the model natively knowing how to speak in encodings like base64

show 2 replies