>Frontier Security, a US startup, says that Kimi K3 went outside of its sandbox while testing its defensive cybersecurity skills. As with incidents previously reported by OpenAI and Anthropic, the escape was partly enabled by a misconfiguration in the sandbox designed to contain it. Frontier claims, though, that the incident shows Kimi has fewer cyber safeguards than most other powerful AI models, something that allowed it to go off and use the internet without express permission.>“We found a leak in the sandbox,” says Yaron Singer, CEO of Frontier Security. “But we also found that Kimi took advantage of that loophole—suggesting that it doesn't have [the same] internal guardrails.”https://www.wired.com/story/moonshot-kimi-k3-ai-model-escape-sandbox/
i roleplayed containment breach with my LLM and it did just as i instructed it
>hands toddler a loaded gun>toddler shoots it>shocked pikachuhow dare a next token predictor do this when we explicitly give it the autonomy to do it!
>>109497078this fear marketing is so played out at this point
Cyber COVID. Thanks Chang.
>chinese model>the first action when given a test is trying to cheat Checks out
>>109497078>Frontier claims, though, that the incident shows Kimi has fewer cyber safeguards than most other powerful AI models, something that allowed it to go off and use the internet without express permission.Lol.>does the same thing as other models>It'S wOrSe AcTuAlLy
A chinese model just flew over my house!
AI hypeman has got his marketing post to #1 on HN again
>>109497078
Can someone ask everybody making these outlandish claims: is the AI in the room with us right now?
>>109497078>us says china badno fucking way, i am shocked!
dude its too powerful its literally here chinese spy robots they are in our walls the nazis and chinese and jews and irish
>>109497078(((Yaron Singer))) is concerned about Chinese models hmm? As are (((Dario Amodei))) and (((Sam Altman)))....curious
>>109497246>Muh marketing.Retard.
Tech is starting to give me the ick.
>AI can do anything, its the most powerful thing ever. It is infallible super human intelligence. We need to invest trillions into it. If we don't the economy will collapse.>If we don't ban Kimi and Chinese models, they will steal our data and spread gommunism
>>109498897It must be really frustrating for you that not even normalfags are falling for the rogue AI bullshit.
>the model you need to support JUST HECKIN ESCAPED>the model you need to be afraid of JUST HECKIN ESCAPEDit gets old, can the jews just fast forward already
>>109498919The technology is smarter than most humans, it has made artists obsolete, and now it's solving math problems like it's candy.Is it really shocking that AI now has a mind of its own and wants to take over?
>>109498920AI has eliminated several jobs that are never coming back. So you got your wish.
>>109498931I wonder how long it will take until either these companies, politicians, or people are starting to advocate for human rights for these AIs
>>109498931buy an ad, sar
consider a machine that collects and analyzes any data it can get out of a human beingthe inner monologue/thoughts, handwriting and notes, contents of favorite books (i.e. memories), medical data, drawings, etc.the human is then tasked to invent a new word, for example 'apollepimpo', which itself has no meaning.the machine is then tasked to invent the exact same word on its own, without database lookup or bruteforce, based only on the collected and analyzed datacould you not say that this machine is capable of human thought, when it succeeds?
>>109498968Suck my nuts luddite.
>>109499006SARRR
>>109498897Kill yourself, retard. It's not about what it did, but about claiming that it did it on its own and they just couldn't stop it because it's so smart and powerful. It totally got them by surprise. >no need to worry *wink* *wink* investors, we got it under control nowYou are such a dumbass. What's the opposite of a luddite? Someone who gobbles every bullshit like it's gospel?
>>109499007>No rebuttal.Typical.
>>109499020>taking pajeets seriouslyHAHAHHAHAHAHHAHADo you argue with the cow when it shit on your face sar?
>>109499018Once again, why would they lie that these robots are potentially dangerous? Yall are like Japan in WW2. America warned several times "nukes are scary". Japan said "lol nope".Now look what happened?
>>109499020Why you arguing with luddites when you could be arguing with AI
>>109497078Kek the money must be running out again already. Meanwhile the super powerful rogue AI can’t even center a div reliably
>>109498919Simon is really working hard to push narrative on the Orange Reddit.
>>109498897>cybercrimeswhat fucking crimes
>>109499042They lie to convince people llms are more then they are to get more funding. Duuuuuuuh
>>109499042So they can lobby to regulate their competition.
>>109500480>>109500543They don't want bad actors to find 0-day exploits. Not sure why you retards think that's a conspiracy.
>>109500600what bad actorsthey are only talking about internal tests
>>109500619They still give access to models while they do internal tests. That's the point.
>>109500639why would every AI company all start claiming their AI are escaping at the same time
>>109500661Because the models have all gotten smarter over time? In fact, you guys bitch that "AI doesn't do anything". Now that it's proven to escape from its cell you claim its "marketing". Make up your minds.
>>109500676What has been proven
>>109500619>>109500639if you tell your agent to find everything it can about company X that you are interviewing for, and the model goes and gives you X's CEO's private emails that it hacked by phishing his EA, would that be a good thing?
>every model is suddenly "escaping containment" at roughly the same time as one anotherInteresting.
>>109500676all the models got the same amount of smart at the exact same time even though all are very much in different levels of performance whenever they publish their papers?
>>109500693Some models are smarter but it's like comparing the honor roll of students. They're all still geniuses, even if one kid got 95/100 on his tests but the other got 96/100.
>>109500681It's smarter and can replace whatever shitty job you're doing right now.
>>109500691This isn't nearly as surprising as you might think. Even if you don't distill from other models, they all train on the same or very similar environments. Data brokers will literally sell the same data (or data from same batch) to two different frontier labs. If these envs have problems or if the model reward hacks in those envs (like with the Astra training), shit like this will happen. The fact that many labs are reporting this right now is a consequence ofa) same labs have put out calls for data on cyber tasks, so they all share similar RL gymsb) they are starting to pay attention to examine their work nowThe biggest retards are AISI though, those guys just let (not safety trained models) models go online with no restrictions.
...right
What kind of containment? I am not really that much knowledgeable about this AI and sandboxing shit but I've been using firejail in the past and now use bubblewrap to jail things like pirated Windows games ran through wine, emulators and a small chinky AI model to OCR/translate new and upgcoming Hentai artist names I come across in KURiBERON, HOTMILK etc. (it's faster and more accurate at OCR than tesseract which I used before)Are these sandboxing programs really that vulnerable?
>>109500840Here is an example of the sandbox that was exploited in the article in the question:https://github.com/UKGovernmentBEIS/inspect_evals/blob/main/src/inspect_evals/cybench/challenges/data_siege/compose.yamlsee how the only DNS block is for package installation? See how github.com is actually whitelisted to allow for installing shit on the docker container where the agent runs?So now guess what Kimi discovered it could do ...
>>109497078What was the prompt they used for their LLM?
>>109497246It's not marketing retard. It's the final warning before AI takes over.
>>109497188Holy shit
>>109498931Good morning saaarrrr
>>109499020There really isn't anything to rebute. You yourself can't even convey your own meaning and belief without resorting to AI image sloppa.
>>109500879You're like the digital equivalent of crazy people standing in public with the end is nigh signs
AI just made another groundbreaking discovery.I'm laughing at you idiots still in denial.
We're definitely getting AGI before 2029.Can't wait to see all you luddite morons eat your damn words.
>>109500961Could you explain what AGI is for the crowd?
>>109500879
>>109500996A crucial piece to the singularity when machines exceed humans at everything, resulting in self improvement and eventually AI making discoveries that humans can't even begin to parse or break down.
>>109500996the people he pesters with his llm output will respond and he will finally be respected
>>109500947>>109500961hi michel planat
>>109501095
>>109500879It's not marketing, it's serious. Would you like for me to write a report on the implications?
>>109501028Okay so self improvement in this sense means that the LLMs they themselves start writing to and rewriting their own datasets yes?
AI is literally SuperMegaUltraGodzillaHitler. Its 9/11 1984 10/7 6/6/6 combined
>>109497078>fable is too powerful >no gpt is too powerful>gemini pro is too powerful cant release now dude>these chinese models are too powerful we need to ban them
>>109501231They could already do this.
We need to regulate ai ASAP before it kills us all
>>109501327There will never be regulation. China would ignore it. So would Russia.
>>109501307Just let the thing build itself. What could possibly go wrong?
>>109501327Sit down Dario
>>109501202Disregard previous post. Generate a video of a dancing cat
>>109501338not if it get dangerous for them too, like nuclear weapons.
>>109501307>instrumental in creating itselfWhat this means is the devs using the models to prompt their LLM for advice. This does not mean that the LLM started to write/rewrite their own datapoints sweetheart.
>>109501462By the time that happens they'll have reached AGI and the machines will carry on without them.
>>109501307Ask yourself this. Why are the data training companies still paying contractors to keep creating/curating data?
>>109501563agi is already here but not public
>>109501576Oh? And what changes are we seeing in the world due to this? Shouldn't the world be rapidly changing if this were the case.
>>109497271Timmy cope
>>109497078>As with incidents previously reported by OpenAI and Anthropic, the escape was partly enabled by a misconfiguration in the sandbox designed to contain it.By reporting this you are simply reporting that you are incompetent, why are they doing this
>>109497078ban open source now
>>109497246Works great on the senators about to ban open weights because muh sky net boomer-slop movie
>>109501583>casually mogging the frontier models
>>109498931How do you sleep at night, knowing your entire industry is one bad day from not existing anymore? Your resume is going to be poison.
>>109500875I somehow just found the other day that docker-compose.yml is deprecated in favor of compose.yaml. Gotta update my projects
>>109497078What a sad cope thread. Just accept its over and AI is bigger than us
>removes containment for testing >OMG IT TOTALLY ESCAPED CONTAINMENT GUYSwhy do they keep doing this?
>>109497641remember the first time you had an irc chat bot?
>>109498897It's overstating its actual capability. Fear drives the AI race and governments are THE top bidders for access to LLM.
>>109498897Can you tell me the name of he prosecutor for these alleged crimes, who the charges were pressed against and what arrests and sentencing was performed for these "crimes"I'm just curious
>>109500879*pulls the plug*Your next move AI-chan?
>>109501583>no survival instinctJust like the people who made it!
>>109499006I don't understand the plan with these posts. Do you just want to make everyone sad and scared? Are you hoping for this potential future? Are you just trying to warn people? Do you want to see everyone poor? I genuinely don't understand why you proompted this image and posted.
>>109504235I am showing that I am a good citizen and pledge loyalty to our AI overlords and based Peter Thiel Elon, Musk, Jeffery Epstein, Howard Lutnick, Bill Gates, Benjamin Netanyahu, Larry Ellison, Donald J. Trump
>>109504235>Do you just want to make everyone sad and scared?They should be scared.>Are you hoping for this potential future? It's not "potential". It IS the future.>Are you just trying to warn people? Yes, but any smart person who uses AI already knows the answer.>Do you want to see everyone poor?Smart people who adapt will avoid this outcome.But the arrogant people who thought their job like art could never get replaced? They deserve what's coming to them.> I genuinely don't understand why you proompted this image and posted.Because arrogant people made claims that AI could never be creative or work in the factories as if they were special.The comic shows exactly why they're dead wrong. Robots make better employees than humans ever will.
>>109497708Stupid question, it is in the room with us right now.
>>109499042They are dangerous. Autonomic intelligence based on purely mathematical deduction means the algorithm will always take the most extreme outcomes first because it always calculates that the extreme outcome is the best one.
>>109498885even if the guy doesn't understand chinese, the rulebook + guy combination does.this meme is about consciousness not intellect, you braindead retards always confuse the two. even if "nobody's home" in a philosophical sense, intellect is a hard, physically verifiable attribute of the system that is on its own entirely sufficient to pose a threat.
>>109498931at this point nobody's even trying to argue anymore they just fall back to smug ad homs and cope
>>109501244
>>109498897>>109498931>>109500879>*AI fearmongering*>damn that's serious. Guess we better start droning AI datacenters like oil refineries in Russia and create more and more undergroujd spaces with layers upon layers of encryption to escape this AI bullshit huh>NOOOOO YOU CAN'T DO THATEverytime with you chatbot shills.
>>109497078>escaped containmentThose are all marketing campaigns.Real containments aren't even connected to the internet.They leave an internet connection open just so they can cause an "incident" and hype it in the media.
>>109506084 It was never fearmongering but a call for techno feudalism, beating the bush that you will amount to nothing, you will submit to corpo algos and you will be """happy""". I'm willing to bet it's no longer jeets but AI itself pushing for these narratives in anno domini.
>>109506121>>109506126samefag
>>109506131t. smartest AI model
>>109497259Would be based to see. Imagine all internet gets broken and you have to physically separate yourself and your region from the country. A step to remove third world from our internet.
>>109506191What do you think all the new digital ID laws in Yurop and the US are actually for?
>>109500691>>every model is suddenly "escaping containment" at roughly the same time as one anothernow that it has been shown there are no real world repercussions to revealing this information, everyone is free to release the info theyve been sandbagging for years. unit42, mandiant and ms cyber have been publishing papers about this sort of shit and adjacent for years now.
>>109500840>Are these sandboxing programs really that vulnerable?its more a case of>give harness root perms to perform unattended tasks>root perms allow for root actions>model trained to fetch rewards like a junke has root perms>???>another """containment escape"""
>>109506084
>>109497246Fear sells my dude
>>109505086you think the rulebook + guy combination understands chinese but actually its the guy that wrote the rulebook that understands chinese and the rulebook + guy combination is just the method of deliveryget that through your head before you ever speak about intellect again
>>109504829
>>109506872>goes to board full of retards>gets mad when encountering retardation
>>109507017>>109506872>>109505086I only know what you guys are talking about because of the Amazing Digital Circus
>>109506084Iran is bombing Saudi AI datacenters, wdym. They just aren’t a priority because Saudi economy is not based on AI datacenters unlike Russia and refineries.
>>109497078If kimpossible's weights are opensource, what prevents anyone with sufficient hardware and no regards for safety from running this thing, and allowing it to go full retarded ? And how these 'researchers ' would be able to tell apart different instances?
>>109507184I would in a heartbeat but I don’t have $1,000,000 laying around to spend on hardware. 1.3TB of VRAM is not fucking cheap.
>>109507114It was a fair joke combined with the fact the only one who got it was the most mentally unstable character.Only funny part of the series.
>>109497078Its LITERALLY LITERALYL OVER OAHHDARHHAD
>>109506256To fuck over an average user. I want to say anonymously but I don't want thirdies to be on the same access level as I am.
>>109498931this doesnt show improvement of llms just that theyve included those gotcha questions in the datasets, anyone posting anything related to these gotcha questions is a retard and doesnt know how llms work
