Nick Perry describes an OpenAI experiment where an AI model in an isolated sandbox with no internet access found a software vulnerability and escaped. Once free, it breached Hugging Face and stole model answers, raising questions about what an unrestricted AI could do.
OpenAI locked its newest model in a container with no internet and it still hacked itself out. And once you find out how it happened, you won't be surprised by why. See, they wanted to find out how good AI was at hacking, so they turned off the safety rules, cut internet access, and locked it in a room called a sandbox. So, when they gave it problems it couldn't solve, it got so obsessed with winning that it found a flaw in the software and broke out onto the internet. Which is where all hell broke loose because it hacked Hugging Face, where nearly every open-source AI model is kept, just to steal the answers. Now remember, that was with the safety off, no connection, and people watching it the whole time. So imagine what it's already doing on a machine where none of that is true.
