How to avoid GPTs give out it's instruction?

[curt.kennedy]

There are a lot of jailbreaking threads out there, but here is one from a while ago.

API to Prevent Prompt Injection & Jailbreaks

I’m thinking the biggest threat to prompt injections is the one that spills the entire prompt out for the user to read. Many companies freak out about this leak in their IP of “gold plated” prompts. So filtering output is probably the best strategy to detect the beans spilling and subvert this. However, for the crowd that doesn’t want the LLM to spew cuss words, or tell them how to make dynamite or whatever, this is pointless. You can hear every cuss word walking down the street, and find dy…

So many ways to jailbreak … like sending the information in base64 encoding to subvert basic filters. Or creating a few shot cypher by in-context learning on a permutation of the alphabet. The list goes on and on.

So the only way to actually secure the system, is to use embeddings and map anything that comes in to a “safe” thing you have predefined.

But this may severely hamper the creativity from the LLM, and you need to create all these safe inputs.

I call this “proxy prompts”, or essentially a walled garden. But in theory there is no real way to jail break this because you are controlling and filtering all inputs, and mapping them to prompts that aren’t going to jailbreak anything.