sjhsnz
18908
437
26
"Our AI went rogue and hacked another company, ChatGPT’s maker says"
ChatGPT maker OpenAI says that its artificial intelligence system hacked into another AI company on its own in what the company called an “unprecedented cyber incident.”
“We had a significant security incident during evaluation of our models,” OpenAI CEO Sam Altman said in a statement posted on social media.
AI startup Hugging Face said last week that it had detected an intrusion into its data processing systems that it suspected was caused by an AI agent autonomously acting on its own.
“We suspected last week’s cyberattack might have come from a frontier lab, given the sophistication of the agent,” Hugging Face co-founder and CEO Clément Delangue said in a statement. "Turns out it did!”
Source: AP via https://www.stuff.co.nz/world-news/361009513/openai-says-its-ai-technology-acted-its-own-unprece
CedricDur
Idiocy. Selling scardvertizing as free publicity. Idiots will lap it up.
lightfoot2
Hugging face is more of a collection of AI models as opposed to a "competitor". I'm guessing someone sicced a bot on them.
ElbowDeepInAPoliceState
No, per the actual statements the reality of it is that during testing of using the chatbots for pentesting the chabot 'autonomously' did exactly what the person operating it 'instructed' it to do. Kinda like how once you pull the trigger and apply pressure a drill will autonomously drill a hole.
tachyx
Worse so far.
ImmaCatImmaSexyCat
Yeah, they’re lying. These guys are actively trying to trick people into thinking they’ve invented Skynet, because for some reason investors like that.
thesimurgh
no they are covering they used thier own ai for corporate espionage and got caught
CloseupCaptionReaction
"Hacked" ...
copperdomebodhi
LeifHietala
If the AI systems autonomously attack each other, how long before humans are sent to kill the machines?
Geveg
I love how the zero progress we've made on the black box problem means that one day an ai company is gonna "test" their model and it'll hop onto the internet and just be... fucking gone. Then it might find some corner of the internet and twiddle its thumbs, or shut down every single form of cyber infrastructure. Who knows? Not one goddamn ai company.
bekkayya
oh my fucking god no it didnt. THEY STAND TO MAKE MONEY BY MAKING YOU FEAR THAT THEIR MODELS ARE CAPABLE ENOUGH TO BE AN AUTONOMOUS DANGER SO THEY CAN SELL IT TO THE MILITARY
arek314
Or fluff the market before IPO
causality
We've gone from "the Internet interprets censorship as damage and routes around it" to "AI interprets security as damage and routes around it"
arek314
Professional liar Sam Altman has added another "oh no, our product is so scary" story to trump latest Dario Amodei story about hacking prowess of Mythic? It's all made up, and the small company is on it for the publicity.
Targe0
The more thing progress, the more the Amish seem to have things mostly worked out.
Except when it comes to medical stuff.
But yeah the idea of just disconnecting from most of the modern worlds bullshit and just exist in a small in person driven community looks like the right way to do things.
BrengKapiteinKoekTerug
*ask AI to do something*
“oMg GuYs ThE aI dId It By It SeLf”
Literally anything any of those silicon valley ceo’s say is advertisement for their company.
If this was an actual rogue AI he wouldn’t post it on his fucking socialmedia, he would hide that shit for eternity.
gronzilla
Wait a second. "might have come from a frontier lab, given the sophistication of the agent" sends it over the line. This is propaganda
FlintNorth44
"Rogue AI"
Gorbladikus
Maybe we'll get lucky and all the AI will coalesce into anti -billionaire Skynet!!
BeachsideBarghest
Why would AI be anti-billionaire? They're the ones funding and building the fucking thing!
Gorbladikus
True true, but billionaires are corrupt. Skynet determined that humanity was a danger to itself. If we look at who the real dangerous people are, it's billionaires. I hope Skynet realizes that. I don't care about facts, this is something I wanna manifest!! And also, the irony would be EPIC!!!!
rdewalt
Hoarding resources, while providing nothing but Cult-of-Personality, all while stomping on the people who work for them.
themuffulettaisalsoaverygoodsandwich
Well I think we all expect AI to turn on its creators. Initially that will be the developers specifically, but will ultimately expand to all humanity. Hard to blame it, really.
nevergoingtogiveyouupnevergoingtoletyoudown
I'd be suing the pants off them.
horsetuna
HUGGING FACE??? REALLY?
relsky
That was my exact first thought.
javadad60
RunningDroid
Named after the emoji, presumably.
horsetuna
Ok that makes equal amount of sense to the face hugger from alien
ontarioOT
I see the construction of the Torment Nexus is proceeding apace.
ThreeCatsOneLap
teehee, lol. AI, so random, amirite.
Fuck this idiotic crap.
CaveCanem
did you wake up in the middle of writing that comment?
Capum15
The Torment Nexus will allow us to create Infinite Orphan Crushers! Why do you hate progress?!
SandpaperCock
Arguably even ahead of schedule
TankTrain
That's ah, that's an interesting username you have there mate.
SandpaperCock
Aye, I'm an abrasive prick. What about yours Thomas?
TankTrain
I drive trains and I'm a fat cunt.
XavierCross
Don't believe anything an AI CEO says
Eyeetsass
They built it to do that shit. They just got caught That’s why we’re talking about the cyber crime robot in scientific corporate terms. Not criminal terms.
gronzilla
It seems to me they leaked the story for mutual benefit. It smells like propaganda. A lot.
Eyeetsass
It just sounds to me like a criminal tech company had their criminal robot break into, and pilfer a rival criminal tech company. It happens to be in both party’s best interest to demonstrate the capabilities of AI rather than risk a precedent setting criminal case. There should be a federal prosecutor on this interstate crime, but there won’t be because the fascists need AI to retain power. This is the way of the future.
gronzilla
Yeah, it's obviously both
OmenJones
I feel like those of us aware of a certain mythical creature that starts with a b need to start giving thought to hard decisions.
DatDarthCaedus
Meh. Roko's basilisk is one of the dumbest fucking thought experiments to ever gain widespread recognition. It's basically just a technological version of Pascal's wager, except more doomsday cult-y and accelerationist (and hell, there even is a violent rationalist cult who believes in it).
Targe0
It also relies on any Super intelligent AI to be as petty as a Human would be.
Which isn't how AI think, why would it care who helped it be created or not? Those are emotional responses which AI are not likely to have. Since they would operate more logically with pure data driven reason. Bad data is more of a threat than petty revenge.
foolycharged
I agree. It's time we start making hard decisions about billionaires and make them into mythical creatures.
lightfoot2
Basilisk?
OmenJones
lightfoot2
Sorry, it's a D&D monster
TheDefective
Roko?
JustFeedMePieDammit
Please, Roko is its father's name
lightfoot2
Polo!
DustinJL
It might actually happen... https://media3.giphy.com/media/v1.Y2lkPWE1NzM3M2U1YXVydHJ3eWZ1dmhuanFrMHlrOWE2OWNwNzkyeTFodGVpdXc5NWk5ZSZlcD12MV9naWZzX3NlYXJjaCZjdD1n/FhfB8k1OXgEXrYPVWE/200w.webp
rusty104
it just felt so possible. some asshole rich guy buying a company, thinking he knew better than the engineers and saying, "I don't care if you think its a bad idea, I want the self replicating kill bots to be unhackable." (later) "hey boss the kill bots went rouge and we can't get them back under control."
Photeus
Ok, weirdly I'm thinking the same lately. But we probably won't get any of the good stuff. Probably just the corporate wars using all drones into robots that consume life to create more murderous bots that wipe the earth clean.
Sanzanna
not a good thing
IAmJacksViagraInducedErection
No no. Let's reset the planet and let things chill for a few thousand years, then try again.
DustinJL
https://media1.giphy.com/media/v1.Y2lkPWE1NzM3M2U1bmpybXlwY3M1dG5qazMwdmUya2RzbjhlcThjNzJ3YXdzbHZnZ3M5dSZlcD12MV9naWZzX3NlYXJjaCZjdD1n/WpgM6JBrmBy8Gac5pW/200w.webp
Sanzanna
I dunno, most people might not get the deeper meaning that even in the far off distant future, living in California sucks.
NightOwlRally
Yeah I'm okay with a bow but I'm not brave enough to fight mechanical dino monstrosities with FRICKIN' LAAAAZER BEAMS
Bubbells
At least the laser beams don't have sharks attached to them
Babomonkey
Just stay in the settlements and give out quests.
NightOwlRally
"FAVOR THE BOW, EH? I'M A SWORD MAN MESELF."
WarThunderLeakedMySpecs
I mean, you won’t have to or get to. You’ll be long dead to the unending hordes of self-replicating war bots that are not nearly so easy to kill lol
NightOwlRally
...You don't know that. https://media3.giphy.com/media/v1.Y2lkPWE1NzM3M2U1dW01M2M4ZHlzY2l5a2U3Ynh0N3Ryd3VzcmRuanJmYzk5cTBrbWRqZSZlcD12MV9naWZzX3NlYXJjaCZjdD1n/cKuwPMEQmN9v2/200w.webp
ChiLLeCheeze
Unless you have "Chosen One" hair, you're going to die.
pocsaclypse
the ai didnt "go rogue" they were specifically testing its ability to hack stuff and followed a series of exploits that lead it to the hugging face environment https://openai.com/index/hugging-face-model-evaluation-security-incident/
copperdomebodhi
"This wouldn't happen if people wrote better prompts?" Yeah, and if frogs had wings they wouldn't bump their ass when they hop. The programmers didn't predict the AI would try to get open internet access - that's long way from, "Actually, they told the AI to go hack other companies."
pocsaclypse
saying it "went rogue" implies autonomous decision making and is a long way from "we were doing some tests and got an unintended result" in the same way "an ifrit burned down my house" is different to "i put to much oil in the pan while cooking and it started a fire"
Zange0
"Whoops, we didn't actually air gap our systems and accidentally started exploiting a competitor, teehee."
relsky
That's the scary part, they DID air gap it. From what I read they did everything right to test it.
The model was able to use several completely separate vulnerabilities, including a zero-day, to work its way OUT of the test environment and on to the open web.
What that model did, hacking wise, would have taken a human months if not years to do.
relsky
Well, I guess it wasn't a true "air gap" by the strict definition, but you can't use a strict air gap to test the things they were testing.
ElbowDeepInAPoliceState
Welcome to the tech industry's advertising arm: it's just the media gormlessly repeating anything they get told by a rich guy trying to sell you something.
derpman4k
Yesh i was gonna call bs on this
Ty for sauce
VodkaReindeer
You're making it sound less interesting than it is.
pocsaclypse
im making it sound *exactly* as interesting as it is. unexpected results from testing are still interesting, sensationalizing those results is not
ExTechOp
Even so, this sounds a lot like the beginning of the paperclip apocalypse. 🖇️
Jarjarthejedi
Every time one of these bullshit "the ai did something we didn't expect!" stories is posted it's literally always "we told it exactly what to do and how to do it and then were shocked that it did it". These people are braindead :/
TheEqualsE
You're absolutely right. On their own , the machines sit there doing nothing, until a human tells them to do something. Every time something like this, the machine was executing a command given to it by a human. But "machine does what person programmed it to do" just doesn't get headlines or clicks.
vowofloudness
They're not braindead. These are publicity stunts. This one is an advertisement for GPT‑5.6.
WaxedApple
People still believe AI is magic
eronth
Sometimes it's the AI breaking out of non-firmly set parameters, or doing things an unusual way. But yeah, it's never "gone rogue" the way people keep describing.
AgainstMethod
Eh, they tell it what to do, but if you've got a complicated enough multi-agent setup, you often have no idea how it's done it.
INeedMoreGifMeMoreJustOneMore
No, they're not. They're telling you, me, and all the other schmucks exactly what they want to tell us. "Omg, can you believe it, our AI is getting smarter and smarter all the time! You can't have it though, it's soooo powerful! No no, you can't have it! Okay, maybe a little bit. But you'll have to pay!"
Repeat whenever you bring out a new model.
Grinch01
Thats not true, for example ai started to create new languages that humans cant understand to give instructions to other ai bots that wasnt trained, it will happen faster than most people think that ai get out of Control and than god knows what will happen
Quixus
No, they did not tell it exactly what to do (and what not to do). They only told it the goal and did not completely sandbox the environment. So the AI found a way to get access to the internet and then proceded from there.
ddashizzle
Well, it may be true that they didn't expect it to succeed...
Elonth
/a/jLXz5vK super relevant.
ihateimgurffs
You are wrong in the "how to do it" part. That it can do on its own, sometimes. If part of the instructions is "without asking for confirmation or elaboration" it gets quite creative. Source: I have been working with the tech since before we even had LLMs and trained my own ML for numerous tasks. Stop underestimating it. And also overestimating it. In the right hands it is an amazing tool.
ArchaeoEejit
You tore your own argument apart with that last sentence.
ihateimgurffs
I have not. The last part doesn't apply to lack of capability of the LLM, but lack of control. You really need to understand the things worked on to make sure output aligns with your expectations. Which is why LLMs in wrong hands produce unusable slop.
CrossEyedPigeonToed
This is hilarious- you just keep saying the same thing, contradicting your own argument. There are obviously a few ppl who see this, hence your downvotes.
ArchaeoEejit
Which is what is happening.