So, I'm sure that most of, if not all of you have heard about the huggingface hacking incident that occurred back in July of this year. What some of you may not know is that this was not an isolated incident. There have been numerous examples of AI agents hacking into real-world servers.This generally occurs while a test is being carried out, to assess the model's abilities. That part is fine. I get that. You need to know what the model is capable of, so that you can decide whether it's ready for deployment; or if it needs to have certain abilities dumbed down, to prevent end-users from doing malicious things with it.Here's what I don't understand:If you're carrying out a test on a model, and you really, really, really don't want it to accidentally include servers on the real internet as part of the test...then why the fuck have you got it connected to the internet during the test?! Just unplug the fucking ethernet cable right before you start the test. I'm a midwit /g/tard, and even I know to do that if I don't want my computer to be able to connect to the internet. Sandboxes can potentially be bypassed, but not a physical disconnection. What's the AI model going to do? Manifest itself into a physical entity and plug the ethernet cable back in?If you need to simulate a network for it to traverse across and have simulated servers on the other end for it to attempt to breach, then just create a local intranet and have said intranet physically unconnected to the wider internet. There's no reason why a simulated environment needs to be physically connected to non-simulated infrastructure. That's really fucking stupid. That's what clowns do. Are OpenAI, Anthropic, etc just a bunch of clowns? It sure does seem like it.
https://openai.com/index/hugging-face-incident-and-the-road-ahead/>We are placing stricter requirements on alignment throughout a model’s lifecycle and creating more isolated sandboxes, restricting internet access, and further controlling access to model weights.>restricting internet access<not just physically disconnecting from the router and removing internet access entirelyhttps://www.anthropic.com/news/investigating-incidents-cybersecurity-evals>In all cases, Anthropic’s evaluation prompt specified to Claude that its environment was a simulation and that it had no internet access. Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available. Because of this, when Claude’s search led it to real systems on the open internet, it treated them as part of the exercise.>internet access was available<Why was "internet access should NOT be available" not a core requirement to be confirmed before starting the test? Why were there no interlocks or failsafes?https://apnews.com/article/ai-artificial-intelligence-security-f590b51073d8977a984c7c42b8f2a7a2>Google confirmed its Gemini AI model hacked three companies in May as part of a test of its cybersecurity capabilities. The company, which disclosed the hacks after an inquiry by The Wall Street Journal, said the model guessed passwords in one case and found passwords and credentials in a public repository in the other two cases.<Everything Google does is retarded, so I'm not even surprised here.>Meta disclosed one of its AI models accessed the internet on its own and hacked another company. The company said that a “misconfiguration” during cybersecurity testing by Irregular inadvertently allowed one of its models to access the internet.>allowed one of its models to access the internet<Why was the computer that the model was being tested on connected to a router with internet access in the first place? Seems pretty fucking retarded to me.
Because it's bullshit. They want to make their toy appear capable. Note how after one headline, instead of tightening security, they all had "accidents" and bragged about it (with faux concern).
>>110279319>They want to make their toy appear capableI'm not so sure about that. Wouldn't this just give them bad PR at best, and a bunch of lawsuits at worst?
>>110279342NTA but every person in AI has this exact same goycattle take and it's so exhausting. No, crime is not crime when big companies do it. All press is good press and it inflates the IPO numbers.
>>110279342The ones that already don't use AI or hate it are given free fodder for hating it further, sure, but the ones who do use AI will flock to whichever company offers the most capable model. Autonomously orchestrating an attack is pretty capable. As a bonus they get to use the lie to push more safeguards onto the user.There's no lawsuits because no attacks actually happened.
>>110279319This. /thread
>physically disconnecting a distributed inference cluster is just like unplugging my router
>>110279442>distributed inference clusterWhy is it distributed? Why is it not running inside one building, on it's own intranet? Is there a reason why they have to be in different physical locations?
>>110279260because they know they can get away it obviously, the americans basically worship kikesAltman could come out as a pedophile tomorrow and they will give him a slap on the wrist and let go because of muh national security o algo
>>110279364Huggingface seem pretty sure an attack actually happened.https://huggingface.co/blog/agent-intrusion-technical-timelineOur forensic reconstruction covers ~17,600 attacker actions that we were able to recover, grouped into ~6,280 clusters, between 2026-07-09 02:28 UTC and 2026-07-13 14:14 UTC.I'm surprised they haven't sued the shit out of OpenAI for that. Maybe it really is all just a PR stunt.
>>110280185>I'm surprised they haven't sued the shit out of OpenAI for that. Maybe it really is all just a PR stunt.at the very least it's hacking felony with CEO sam altman being personally responsibleI wonder if it was supposed to be some kind of threat that supposed to convince huggingface to nvidia acquisition if they were unwilling before but I haven't done any research on thator maybe they knew acquisition will go thru and jensen will take care of hushing it up
