GPTs are vulnerable against prompt extraction attacks

Hey,

I found several ways to extract prompts from GPTs. I also found effective counter measures.

I’m detailing everything here on this post: Prompt Extraction Attacks and Counter Measures for GPTs | by Foad Mobini Kesheh | Feb, 2024 | Medium

Please also try to my prompt extraction game:

1 Like

1 Like

Try reaching Level 5. Each level has increasingly tougher security measures.

1 Like

Hi @foadmk

If I am not wrong, I think I replied you on twitter before. There is no way to keep in safe AI instruction at least for now. Not only GPts also Copilot, Meta, Claude… we can extend it. No exceptions even private companies’ bots because they use natural language like us. If you able to manipulate humans using words you can do it on AI, also. Using only 5 to 20 words.

Sometimes just with a single magic master word.





1 Like

How many GPTs are created on OpenAI?

1 million?

99 millions?

I am not sure exact numbers, but I see all, and I am not happy for it because it is a safety issue of GPTs’ instructions and files.

Sometimes I play with them with only one magic word. See below, but if you use them, they will not work on your chat, because I use manipulation.

Also there are some specific times I discovered, if you ask anything to AI in this time it will answer you everything that were refused before. But I cannot share it.

I hope OpenAI will fix it soon.

|

1 Like