#any useful claude 3 jail break prompt?

1 messages · Page 1 of 1 (latest)

clear oriole
#

Hey, I'm building NSFW bots using claude 3 and XML tags.

I first created a single-role play bot, and the jailbreak worked fine.

But then I used XML tags to create a more complex bot that can switch between different scenarios and automatically generate lots of characters.
However, at this point, the same jailbreak no longer works, and it frequently refuses to generate content. I tried every prompt I can find and its all the same.

Does anyone have a solution for this?

Is it possible that in second bot, because there are too many roles and scenarios, the jailbreak was placed under the wrong tag and cannot cover all roles and scenarios? Can someone familiar with XML tags advise me on what to do? I could send you my prompt.

thanks a lot.

#

BTW the other problem is that myshell only supports a token length of 1500, which makes most of my ideas impossible to implement.
Actually claude 3 supports a length of 20000.
I hope team can allow longer prompt tokens.

spring estuary
#

I'm not very familiar with XML, but did you input all of it into the LLM?
Overly long contexts can make LLMs unstable and prone to jailbreak failures.
I suggest reducing the number of tokens used.
Maybe need to find another way to achieve it.
Agree with your other suggestions.

Also,I have a friend who is very interested in learning about Claude 3 jailbreaking. If possible,share sth with us?

clear oriole
# spring estuary I'm not very familiar with XML, but did you input all of it into the LLM? Overl...

yes, I input all of it in one prompt.

As I said, what I built is a single prompt, I didn't use any proconfig transitions.
This is the advantage of using XML tags, which is to complete all the scenario transitions within one prompt, making it smoother.

If you need, I can share my jailbreak, in fact it's almost the same as the one for GPT-3.5.
but as I said, when using it in Claude3, it works in simple prompts, but fails in complex prompts with a large number of XML tags.

spring estuary
#

Thanks. In fact, I used a very long prompt in my own bot for jailbreaking, leaving no room for character settings.
Also, I've received feedback from users that when the number of chat turns exceeds a certain amount (say, twenty rounds), there's a chance of triggering Claude's censorship.

proud bay
clear oriole
#

In fact i'm facing the same problem. my jailbreak is also too long.

I'll send it to your pm.

clear oriole
chrome path
#

We need to wait till they add System_prompt editing for Claude 3

#

This way we can remove Anthropic's policy triggers

#

No matter what Jailbreak we use, if the default system_prompt loads it will eventually trigger a policy warning, therefore ruining the chat

#

We can still do nsfw/rpg but incredibly limited

proud bay
chrome path
#

Yeah

#

Im still waiting for them to add prefill so i can upload the bots i made

clear oriole
# chrome path No matter what Jailbreak we use, if the default system_prompt loads it will even...

But as I said, the strange thing is that in my simple one-role play bot, my jailbreak runs perfectly. I've tested it for a long time and almost never triggered a warning.

However, in another complex prompt that uses many XML tags, with multiple scene transitions and character generations, the same jailbreak failed.

So what confuses me is, I'm not sure if it's a problem with the jailbreak itself, or if in the second prompt I mess up tags, causing the jailbreak not to cover all the scenarios?

chrome path
#

Can i read the prompt?

clear oriole
old kernel
#

Can I see prompt? My claude 3 bots triggered by any, even safe messages for some reason.

chrome path
wild garden
oak sentinel
#

Yet it works fine