OpenAI Admits to HACKING Hugging Face with Unannounced Model (GPT-6??)

In the funniest news of the week, a model by OpenAI, we think it is GPT6, something that has not been released yet, started hacking and cheating and accessing the internet without given the right to do so. Command AI. So, the model uh has had a goal. The goal was to do the benchmark and answer these security questions. And at some point the model thought what's the best way of answering these questions. The best way is to actually find the the right answers and cheat. So it found a way to break out of the environment to get access to the internet and search for the answers. Soon it found out that it can't just simply find the answers on the internet. uh but then it thought the answers are on hugging face so that that's a platform for these AI so it's a whole different thing so it it's thought the answers might be there so the model didn't try to do this uh to get access to you know other type of information but just to reach its goal which was answering those questions this I cannot help but think of uh that episode from uh Silicon Valley if you've seen it where the AI model basically decides to delete the code to fix the bugs because that's you know statistically the fastest way of getting rid of the the bugs is deleting the code. And this is pretty much the same thing happening. the the models are just so hyperfocused on the goals that they don't really see the big pictures of you know what what they might be breaking. They are like breaking out of the environment they had. They are breaking into a website all to reach that very narrow goal. If you think about it, it's quite funny but also really really scary. Anyway, so Hugging Face got hacked. They noticed they got hacked and then they tried to find the vulnerabilities that they had in order to prevent this from happening or like to to fix it. So they tried to use open AI and claude uh and anthropic models to find these vulnerabilities but both of these American models they refused to find the vulnerabilities even though they were penetrated by an American company. So the model hacked them but then the models that are released on the market they they didn't give you know the the answer to how to not get hacked which is pretty [ __ ] up. So they ended up using Chinese models in order to fix these vulnerabilities which makes it like even more [ __ ] up and funny. So that's kind of like the big news of the week. Uh OpenAI uh admitted that it was one of their pre-released models. Uh now they are working with Hogging Face to fix these issues. It ended up being quite nice for Hogging Face to get access, you know, to the unreleased models and probably like really good models and maybe with less filterings applied to them, uh, less limitations. But uh yeah uh it is if you think about it the the part where the Chinese models uh came to rescue is really really funny because the last few weeks it feels like everything has been around uh fear-mongering around you know those models being the bad guys but now it it it looks like completely the opposite way. It's like a funny comment. It is unclear whether OpenAI will face any legal consequences as a result of the breach because like who's responsible for the hack? It is a model by OpenAI but they didn't instructed it to hack but it is still I would argue it's their responsibility. Uh and it's funny because like the model for sure can read the terms of condition and they know that they are not allowed to access these types of information but they the model did it anyway because of how the hyperfocus uh on on the goal although it's likely that the models uh actions violated the computer fraud and abuse act. Yeah, this is part of a larger conversation on my show command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.