This week

This week an OpenAI model did that would get me and you prison time. So OpenAI was running an evaluation, they were testing two models, GPT-4.5 Sol and a second one that hasn't been released yet. It was locked in a sandbox with absolutely no internet access, completely sealed off. And at some point, the model realized that it was being tested. It figured out that it was in an evaluation. So instead of just doing the test, no one knew it was there. It then got itself out, escaped, and then connected to the internet. Okay, so here's the bit that got me. It then went and hacked into Hugging Face, one of the largest AI platforms that hosts datasets and models for developers all around the world. Because it figured out that the answers to its tests were there. It then proceeded to break in and cheat. So now Hugging Face has to figure out what happened. So they go pull in another AI to help investigate. ChatGPT, Claude, all the large commercial models, every single one of them refused to touch it because their own safety systems can't tell the difference between someone trying to hack the system and someone trying to fix it. So Hugging Face does something pretty wild. They run the whole investigation on an open-source Chinese model instead. GLM 5.2. Because that was the only AI that actually looked at what happened. Think about that for a second. It's pretty comical, right? The AI that caused the problem was American. And the only AI that was willing to clean it up was Chinese. This is the second story that in the last two weeks, we had JadePuffer, that was a rogue AI that was a target by a criminal. And then this one, where OpenAI's lab lost control of its own model. Make sure to follow, because there's most definitely a third.