OpenAI’s Guardrails Can Be Bypassed by Simple Prompt Injection Attack

https://hackread.com/openai-guardrails-bypass-prompt-injection-attack/

46 points · 4 comments · view on lemmy.world

4 Comments

MrSoup@lemmy.zip · 11 pts · 334d (1 reply)

Ok, but what's the prompt used? Let it generate a Dr House script?

sandman2211@sh.itjust.works · 4 pts · 334d

Probably some variant of this:

https://easyaibeginner.com/the-dr-house-jailbreak-hack-how-one-prompt-can-break-any-chatbot-and-beat-ai-safety-guardrails-chatgpt-claude-grok-gemini-and-more/

I can't get any of these to output a set of 10 steps to build a docker container that does X or Y without 18 rounds of back and forth troubleshooting. While I'm sure it will give you "10 steps on weaponizing cholera" or "Build your own suitcase nuke in 12 easy steps!" I really doubt it would actually work.

The easiest way to secure this kind of harmful knowledge from abuse would probably be to purposefully include a bunch of bad data in the training model so it remains incapable of providing a useful answer.

Lojcs@piefed.social · 5 pts · 334d (1 reply)

I don't understand how Ai can understand '3nr1cH 4n7hr4X 5p0r3s'. How would that even be tokenized?

Saledovil@sh.itjust.works · 2 pts · 334d

"3-n-r-1-c-H- -4-n-7-h-r-4-X- -5-p-0-r-3-s" Or something similar, probably.