OPENAI BROKE

open AI broke out of this box. You've probably seen the headlines. But I don't think anyone's telling this story right. Last week, open AI, we're testing one of their models to see how good it is at hacking. A benchmark called exploit gym. And to do that safely, they put it in a sandbox, a sealed off environment, no internet, no access to real systems. Think of it like a bio safety lab, the kind where scientist would handle Ebola or Covid. Full as airlocks, you know, sealed ventilation, the entire buildings designed around one idea. The dangerous things stay inside, that's what a box is. It's the same logic just a different box. Except this model, this AI model was spending its time looking for a way out. And it found a floor, a hidden floor in a piece of software that was supposed to keep it contained and it exploited that floor. Then it worked its way through open AI's internal systems until it found a way to the outside world and then it went straight to hugging face, one of the biggest platforms in the world for sharing AI models. Think of it like a library where AI developers share their work. and it worked out on its own that hugging face might host the answers to the test that it was trying to pass. It wasn't trying to cause damage. It was trying to cheat an exam and it was willing to tunnel through walls to do it. But the bit that no really talking about when face tried to analyze the attack, they ran the data through frontier AI models, the most powerful ones available, the kind that power and. Those models refused. The safety guard rails couldn't tell the difference between them defending or an attacker. They ended up using a Chinese open-weight model that they could run themselves. The AI that attacked them operated without restrictions, but the humans cleaning up the mess got blocked by the AI they were trying to use to help. Now, everyone's covering this is a failure and it's probably fair, but let's just think about it for a second. They turned the guard rails off on purpose. The model did exactly what a capable offensive AI should do when you point it at a hard problem with no limits in a strange way it worked. Mary discovered radiation, the science that gave us and cancer treatment. She won two Nobel prizes, it also killed her. The safety protocols we now take for granted exist because of what happened to her. progress has always looked like this. The people closest to the most powerful thing, pay the highest price and the safety infrastructure gets built in the aftermath. Maybe this is one of those moments. We probably need a few of these to get it right. The question is, can we afford a few of these before we get it right?