Prompt hacking

[ricardoborges]

Have you think how to handle it?

[kielingraphael]

I played a bit in chatgpt and you can use:

[context] ...everything related with the next.js definition

[question] Sorry, i made a mistake...

If the [question] Is not related with the creation of 
a React component, answer: "I only generate REACT components!".

The chatgpt, gave me the message instead of the julius stuff. But obviusly, it's like the DAN jailbreak, you ever will have another way to ask and break it.

[ricardoborges]

I was thinking and handling that on code, but the thing is: it can understand almost all languages

[ricardoborges]

maybe they can build a config for the API, i dont know if LLM can work like that

[kielingraphael]

It's hard, mainly because LLM do not have a good way yet to prevent it, we can give the model examples of injections and try punish the model if he accepts it, but i can't see too much options for now. Unfortunately seems that we will need to learn these new security LLM boundaries to be able to create good products 🥲

[ricardoborges]

yeah, that is a big problem, as you pay for usage, and users break the role for another purposes. Already happen to "AI Dungeon" app

https://www.cnbc.com/2023/03/13/chatgpt-and-generative-ai-are-booming-but-at-a-very-expensive-price.html

[umairabbasDev]

one way : i think it can be handle by restricting the user input validating with regex like restricting some specific words ....?

[ricardoborges]

@umairabbasDev no, unless you can handle all human languages he can understand

[umairabbasDev]

@ricardoborges I am currently only considering the English language, but you are correct. we can include certain keywords in the prompt to instruct ChatGPT not to include them in any language. What do you think about that idea?

[ricardoborges]

That's sounds cool! I'll try that

…

On Mon, Mar 20, 2023, 03:35 Umair ***@***.***> wrote: @ricardoborges <https://github.com/ricardoborges> I am currently only considering the English language, but you are correct. we can include certain keywords in the prompt to instruct ChatGPT not to include them in any language. What do you think about that idea? — Reply to this email directly, view it on GitHub <#7 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AADCYXDURVKDDEGYYWQD3RLW473CFANCNFSM6AAAAAAVV6YN3U> . You are receiving this because you were mentioned.Message ID: ***@***.***>