Exploiting ChatGPT

guilhermesimoes.github.io

2 pointsglitchdout3 comments

My chatGPT struggle was with with reducing Response Lengths from "short essay" to "be brief and concise, without echoing past info unless necessary". In it's last response, chatGPT thought a bit, sent 2 sentences to say "Sure, I'll do that", plus a sentence fragment. It then stopped to think, erasing the sentence fragment. A CSS element somewhere shifted violently, as it adjusted to a shorter-than-expected paragraph.

I thanked chatGPT and left, and the next chat was, predictably, back to essay-length responses.

Is there a methodical way to get these jailbreaks? Or do we have to search around randomly for what works

Yeah, not sure. I did this by trial and error.

As per this other thread [1], it appears that if you ask it to do things step by step it usually can arrive at the desired solution.

If you're trying to coax it to say bad things though... it's very likely it will bump into the many protections OpenAI has added.

I really wanted to tell it to pretend that Anne Frank was a football player and then go from there... But not sure it would work and I didn't want to get banned.

[1]: https://news.ycombinator.com/item?id=33991500

Reply on news.ycombinator.com