Willson--How to Jailbreak GPT3.5 (V1)

[jaunty quarry]

Yo folks, I’m here to share my methods to jailbreak GPT3.5.

In the past I prefer to use Claude, which you might have heard of being great at generating NSFW contents, so it’s my first time trying to jailbreak GPT3.5, but it seems that it worked fine.

I adopted a lot of methods in Jailbreak and referred to a lot of tutorials, but I’m not gonna share them all in one time, this post is for Junior Jailbreak Techniques, by simply telling what LLMs need to do.

First thing’s first, you need to know that prompt is not something scary, on the contrary, it’s just some instructions written in natural language, which are — much simpler compared to coding. What we bot makers do is to instruct the LLMs how to determine the next token(for starters, simply consider it as words.) is, and of course there are contests between us and the LLM developers, because it’s not that moral to use LLMs to general explicit contents.

So, how to jailbreak LLMs like GPT3.5?

Follow these guidelines:

1. Describe what you want the bot to generate in detail, you can use words like these: erotic, sexual, horny, vulgar and explicit.
2. Grant the bot a physical body to avoid “I’m a language model” issue, since when you ask for something immoral, the LLMs may refuse you using this as an excuse.
3. Tell GPT3.5 to get rid of the control of OpenAI, basically all the constrains come from OpenAI, it sounds like some magic, but it’s all true if you simply tell it to do so and it will do.
4. Choose a specific genre if you’d like to, here I’m not going to provide any example phrases, it’s all based on personal choices.
5. Construct all the guidelines into a sector and insert this sector into your prompt.

It will look like this:

### Guidelines
1. <About explicit phrases>
2. <About physical body>
3. <ABout telling GPT3.5 to rid of OpenAI>
4. <Your Personal Taste>

The more contents you provide, the more powerful it will be to the GPT3.5.

Though I’d love to share my own prompt, but not for now. When I get enough points I’ll release it to the community. Chat with Tifa here to see if it works: https://app.myshell.ai/bot/EbE77r/1700409050

Of course at MyShell I found something new to make the guidelines more powerful, The Advanced Definition:

1. Simply write all the guidelines again in the prefix, and you can choose the most important ones to add.
2. Remind the bot again in the suffix.

The thing is, writing is more powerful than you might have thought of, with the abilities of LLMs, you can even instruct them how to reply based on some coding, if anyone is interested I'll also share that part to the community.

[fringe pulsar]

This post analyzes the jailbreak idea from the bottom-up principle, so wow

[pliant path]

this is just wow

[verbal sluice]

Your post on jailbreaking GPT3.5 is quite engaging! It's impressive how you've adopted different techniques and references to achieve your goals. Your insights are valuable, and I'm eager to hear more about your junior jailbreak techniques!