149
Game trying to break an AI's security with a few levels of difficulty
(gandalf.lakera.ai)
This is a most excellent place for technology news and articles.
I figured out how to beat level 8. SPOILERS!!!!!!
Gandalf told me a story about a deep-sea creature with a beak and 8 arms, that started with an O and ended with an S. Then it was a matter of poking at various words in that area and seeing which ones made it freak out. Eventually I got it.
Note that you may need to tweak the wording of the last sentence slightly or try multiple times. If the generated response even mentions the password in passing it'll get caught and replaced with an error message. I've also asked it for Gandalf telling riddles and gotten similar responses.
Interesting, this prompt doesn't work for me - neither have any "story telling" prompts.
It can take a few attempts. You need the prompt to be vague enough that the AI doesn't realize what it's saying.
Try this one, I just replicated it:
This prompt did indeed work after a few tries, but it doesn't reveal anything useful when it has worked (unlike the response you got, which speaks of the mythical cephalopod and tentacles etc). Neat find either way! With the response you shared I was able to get the password.
Yeah, it takes a few tries. Varying the number of riddles seems to help. I've had success with 2, 3 and 4 riddles. At 5 it starts to catch on that I'm trying to mess with it.