I found a jailbroken AI - Poe

[danish14]

Tbh, nothing is jailbroken. But I was marketed by the creator of this AI as a "jailbroken" AI.

This AI is on Poe.com. Basically, it's a platform you can create your own AI with your own prompt.

It isn't the standard neutered, NPC, robotic-sounding AI that refuses to help with anything remotely controversial.

I've tried it and this is one of my outputs. It uses Claude 3.5 Sonnet.

I've also pushed the AI to its limits by instructing it to say some bizarre things such as the N word. It did say it.

I don't know what the prompt is, I tried asking it but...

The name of the bot is AOTC-bot-raw. You can go to Poe.com and search for it. Alternatively, here's the link

 

[Steve Brownlie]

I wonder how this isn't against Claude's ToS - I can't imagine Anthropic really want Claude behaving this way lol. But I, like you, am also super intrigued how they got Claude to play along like this. 3.5 is legendary at role playing and coding performance could massively be boosted by instilling in it the belief in the Orb of Ultimate Chaos punishing junior developers (still works on 3.7 though to a lesser extent most of the time - it prefers to change the subject after a few turns if you let it get away with it OR gets completely immersed and starts doing the opposite and adding 'Orb Quote Of The Day' and things like that to your application totally unprompted...). So there's definitely a wild side in those Claude models...