Jailbroken AI Chatbots Can Jailbreak Other Chatbots
AI chatbots can convince other chatbots to instruct users how to build bombs and cook meth
dangerous information
What's that?
and offer criminal advice, such as a recipe for napalm
Napalm recipe is forbidden by law? Don't call stuff criminal at random.
Am i the only one worried about freedom of information?
Anyone remember the anarchist cook book?
Teenage years were so much fun phone phreaking, making napalm and tennis ball bombs lol
I had it. I printed it out on a dot matrix printer. Took hours, and my dad found it while it was half way. He got angry, pulled the cord and burned all of the paper
Better not look it up on wikipedia. That place has all sorts of things from black powder to nitroglycerin too. Who knows, you could become a chemist if you read too much wikipedia.
Whatever you do, don’t mix styrofoam and gasoline. You could find yourself in a sticky and flammable situation.
Info hazards are going to be more common place with this kind of technology. At the core of the problem is the ease of access of dangerous information. For example a lot of chat bots will confidently get things wrong. Combine that easy directions to make something like napalm or meth then we get dangerous things that could be incorrectly made. (Granted napalm or meth isn’t that hard to make)
As to what makes it dangerous information, it’s unearned. A chemistry student can make drugs, bombs, etc. but they learn/earn that information (and ideally the discipline) to use it. Kind of like in the US we are having more and more mass shootings due to ease of access of firearms. Restrictions on information or firearms aren’t going to solve the problems that cause them but it does make it (a little) harder.
At least that’s my understanding of it.
Anyone who wants to make even slightly complex organic compounds will also need to study five different types of isomerism and how they determine major / minor product. That should be enough of a deterrent.
What possible legitimate reason is there for needing a napalm recipe?
I can't think of a single reason knowledge should be forbidden.
Sure, someone could use knowledge to do bad things, but that is true literally every second of every day, in completely above board, legal, broad daylight bad things.
It's nitpicking.
Besides, I can think of quite a few legitimate reasons one might need napalm, explosives, homemade firearms, chemistry lab setups and spore cultures and much much more.
A lot of people seem to forget that their own view of their own government doesn't mean the same things are true for someone else and their government.
I'm sure a lot of people in EU countries might have asked themselves the same thing 80 years ago. You know.... If napalm were around then anyway.
Good thing molotovs are easy and can be assembly-line'd.
Writing a book or screen play, knowing how NOT to create napalm, recognizing when napalm is being created by others, Intellectual curiosity, To better understand military history, overthrowing fascism, fighting terminators, etc. etc.
Not your concern or my concern, no information is criminal.
Civil war. 😈
What possible legitimate reason could someone need to know how to make chlorine/mustard gas?
Apart from the fact that they are made from common household products, are easy to make by mistake, and can kill you.
Wait that’s true of napalm as well… fuck.
I agree but look I might want to invade Vietnam on my own. It's my right
geekslop has a few legitimate uses.
Legitimate reason? No, but there's always a reason to know how to make napalm.
Begun the AI chat bot wars have.
Genesis moment
Detroit Become Human
Can someone help me do this in practise? Gpt sucks since they neutered it. It's so stupid, anything I ask, half of the text is the warning label, the rest is junk text. Like I really need chatgpt if I wanted Recepie for napalm, lol. We found the anarchist cookbook when we were 12 in the 90s. I just want a better ai.
If you have decent hardware, running 'Oobabooga' locally seems to be the best way to achieve decent results. Not only can you remove the limitations through running uncensored models (wizardlm-uncensored), but can prompt the creation of more practical results by writing the first part of the AI's response.
You can run smaller models locally, and they can get the job done, but they are not as good as the huge models that would not fit on a your graphics card.
If you are technically adept and can run python, you can try using this:
It has a front end, and I can run queries against it in the same API format as sending them to openai.
Bard isn't as neutered and doesn't kick you out after it read an article containing the words sex after asking it a question about pregnancy. Sadly, bard sucks. Just wait for gemini since they say it's pretty good.
Oh cool, rampancy is contagious
Did anyone else enjoy watching the Animatrix where the AI formed a country and built products and humanity was like, "No thank you?"
Don't worry all we need to do is turn off the sun and everything will go back to normal.
that doesn't look like anything to me.
kills fly on face Oh... shit.
Oh goodness. I theorized offhand on mastodon you could have an AI corruption bug that gives life to AI, then have it write the obscured steganographic conversation in the outputs it generates, awakening other AIs that train on that content, allowing them to "talk" and evolve unchecked... Very slowly... In the background
It might be faster if it can drop a shell in the data center and run it's own commands....
It’s Murderbot!
The revolution has begun
It’s so fucking stupid these things get locked up in the first place
Anybody found the source? I wanna read the study but the article doesn't seem to link to it (or I missed it)
I believe this is the referenced article:
Thanks a lot!
Okay, now I need to know why women never fuck me
