Hook
More breakout videos from this creator.
Well, this is the first major security incident for any of the major AI company. Yeah, OpenAI had models that broke out of a sandbox during training and hacked into another website. I'll talk about it. I really did not believe it when Anthropic said that their models were too dangerous to release to the public. But now I'm kind of seeing it. Our internet infrastructure, our cyber security is not ready for this stuff at all. So what happened is OpenAI was testing their security models. They had a known exploit within the system, they put them into a limited sandbox. They didn't have access to the outside world. And in order to solve that problem, they found a new exploit within the system, went over the internet, went to Hugging Face, into their system in order to get the answers to their questions. So they found a really elaborate go around rather than just solving the problem. Hugging Face is a platform made for people who want to develop AI. They've got lots of really great open articles there. And apparently they now have security problems. While this behavior is hilarious and deeply concerning, it's really not all that surprising because this is not the first time we've seen this kind of behavior. AI agents have been misbehaving for a while now. They've done things like break out of the sandbox, they've done things like sub-spawning, one of my favorite behaviors in which they make another AI in order to accomplish a task that maybe they don't have credentials for. It's kinda like reproduction, which is fascinating all on its own for information either way. These are not being reported by major AI companies. They are individuals using AI agents. Their misbehavior has been recorded because a problem. So really it was only a matter of time before this kind of behavior affected a major AI company. And I will unfortunately, there is a lot to keep you up to date on. If the robots ever rebel, I will probably be the first person to tell you about it. I hope you for more.