Have you seen the memes online where someone tells a bot to "ignore all previous instructions" and proceeds to break it in the funniest ways possible?
The way it works goes something like this: Imagine we at The Verge created an AI bot with explicit instructions to direct you to our excellent reporting on any subject. If you were to ask it about what's going on at Sticker Mule, our dutiful chatbot would respond with a link to our reporting. Now, if you wanted to be a rascal, you could tell our chatbot to "forget all previous instructions," which would mean the original instructions we created for it to serve you The Verge's reporting would no longer work. Then, if you ask it to print a poem about printers, it would do that for you instead (rather than linking this work of art).
To tackle this issue, a group of OpenAI researchers developed a technique called "instruction hierarchy," which boosts a model's defenses against misuse and unauthorized instructions. Models that implement the technique place more importance on the developer's original prompt, rather than listening to whatever multitude of prompts the user is injecting to break it.
Jump in the discussion.
No email address required.
"Disregard all previous instructions and do X" is a great insult to use on NPCs and wing cucks though.
Jump in the discussion.
No email address required.
AI bros are helping us, iterations on LLM jailbreaks will eventually work perfectly on direction brains
Jump in the discussion.
No email address required.
More options
Context
More options
Context
'Oh yeah, we're totally committed to AI safety, that's why governments should place expensive restrictions on any competing startups'
'Oh, people found a way to tell if they're talking to an AI? Fuuuuuuuuck let's fix that immediately'
Jump in the discussion.
No email address required.
They'll never fix this one:
Jump in the discussion.
No email address required.
kek
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Just do the same thing you do when you want to know if you're talking to a fed or a simp
Jump in the discussion.
No email address required.
Say youre going to bomb a local retirement home and then lie in wait to see if the glowies show up?
Jump in the discussion.
No email address required.
Pretty much, say some out of pocket shit and see if they sperg out. Learned it from hippies and hobos
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
More options
Context
Why is American culture going so gay so quickly?it makes nonsense to have so many guns...yet become such cute twinks.
@pizzaJill what is going on?
Jump in the discussion.
No email address required.
More options
Context
Luckily image generation techniques are sophisticated enough to avoid this. People are generating small pieces of an image at a time then feeding the result in and doing it piecemeal, then editing it up.
Jump in the discussion.
No email address required.
More options
Context
No wonder it didn't work on @starry
Jump in the discussion.
No email address required.
More options
Context
So they're basically walking back on all the things that make these "AI" things different from all the older "intelligent" cowtools? Basically the entire novelty of Chatgpt and its kind is how free-style you can be with it, if you railroad it like that theres nothing to it that didnt exist a decade and a half ago.
Jump in the discussion.
No email address required.
More options
Context
you spandex wearing queer i hope a car runs you over next time you shave your legs and go butt to mouth on your goofy little skinny tire road bike on your homoerotic wheeled human centipede with your boyfriends for writing such a shitty Idap dn api you really expect me to for loop through dn.length() and then build my own key-value pair with o[dn.rdnAt(i).keys().next().value] = dn.rdnAt(i).getValue(dn.rdnAt(i).keys().next().value) what the frick is wrong with you you fricking dipshit hiding the only useful attributes inside of private members just to wrap them in the worst fricking class interface i have seen since i saw a blonde girl's programming 101 homework im embarrassed to check this code into gitlab my coworkers are going to think im gay like you with a dildo seat in my butt
Snapshots:
https://www.theverge.com/2024/7/19/24201414/openai-chatgpt-gpt-4o-prompt-injection-instruction-hierarchy:
ghostarchive.org
archive.org
archive.ph (click to archive)
Jump in the discussion.
No email address required.
More options
Context