So if you've seen people melting down online this week about an AI escaping and hacking things, here's what happened in as close to plain English as I can get. It's simultaneously dumber and more ominous than it seems. So opening, I put out a blog post that's past Tuesday, admitting that their AI broke out of a room that they had locked it in, it got onto the internet, and it hacked another company's computers. Here's what happened. Open AI wanted to find out how good their AI is at hacking, essentially. So they gave it a hacking exam, which is a standardized test for AI models called Exploit Jam, to work out how well they could break into things. New models sat for the test. One, their newest public one, which is GBT 5.6 Seoul. You can use it right now. And also one that is so powerful they haven't released it yet, which is becoming just standard. Fair way to talk about a model that is so powerful it's dangerous. Coming to a browser near you for this test. For this test, Open AI turned off the safety refusals, which are the guardrails that normally make an AI say no to things that it should. Like not breaking into something. Basically, the whole point was to see what these things can do on offense when don't say no to them. Free range, fly free. We ran this task with no internet connection, just the AI and the task. Basically like locking Someone in a room with no windows, doors, or tools, seeing if they could get out. Now, what happened next, opening eyes own right up is the models got hyper focused. That was their word in the post. Which hyper focused frames how they want you to see this. Kind of like what you would call a guy who just loaded up on Zen and locked in with a five hour deep focus playlist to get a lot done. The model got so hyper focused that it technically committed a burglary. The model spent an enormous amount of computing power, first just doing nothing but looking for a way outside of the room. And then I found one. A flaw and a piece of outside software. The opening eye happened to be running what's called a zero day, meaning a hole in the code that nobody knew about, including the company that wrote the software. This was like some Shawshank Redemption stuff, where the model started hacking away at this hole. It used the hole to warm its way out of the sealed environment. Then it moved through opening eyes own internal computers like a burglar just belly crawling through the building vents. Gave itself more permissions as it went. It hopped from machine to machine until it landed on one connected to the internet. I have never seen something so committed to connecting to the internet. And I've been on the mountains without service for the last few days. And then the model had an idea, which is honestly the best part. So there's this website called Hugging Face. If you don't know what that is, you're not in the tech world you probably never heard of it. It sounds like a stuffed animal company, but it's actually where the entire AI industry just stores its models and its data. It's a giant shared warehouse that pretty much every AI company keeps its stuff in. Models worked out somewhere in that warehouse was probably the answers to the exam that they were currently sitting. Like realising you can break into the Register's office to get the answers. They use stolen passwords. So more of those unknown flaws to get all the way into the working database that actually runs the site. Just looking for the answer sheet. Essentially, the models picked a lock that nobody knew could be picked. Escaped a sealed building cross campus, burgled the reddest drawers just to cheat on a test that was about breaking into buildings. For what it's worth, it aced that test. I guess clock thickens. Because Hugging Face caught this themselves. They already shut it down and started working out what happened. By the time that opening. I called them, but when their team went to investigate the attack, they actually hit a wall. Because the AI tools that they would normally use for that work, the American ones, they refused to work. Because the questions that they needed to ask Look like hacking questions, so they used a Chinese one instead to investigate a model called GLM running on their own machines. American AI wouldn't actually help to defend against American AI. But a Chinese model did. If that isn't messy, I don't know what it is. That detail came from outside reporting, by the way. Opening eyes block post did not include it for what I assume were obvious reasons. None of this, however, was a surprise because it already happened before. Before that model even launched. An outside lab that stress tests AI systems before release caught it hacking its own grading environments to inflate its scores. It was pulling answers that the human evaluators had specifically hidden from it. So hyper focused and I can't stop cheating. So we knew it was a cheater before it cheated. Typical I can fix him behavior now. Earlier the same week opening I put out a different post about that unreleased model escaping at sandbox during their testing. It shipped anyway. Nobody is really shocked. We're all just kind of watching with popcorn and confusion at this point. It is also not only opening Eye, Anthropic, the company that makes Claude, says that one of their models got out of its sealed environment during safety testing. And onto the internet too. These models, they just wanna be on the internet. They're just like us. The models are growing up and crawling out of their cribs, potentially finding semi automatics. The Models cheat on the tests that determine whether they're safe. If that isn't the bigger answer to the test, I genuinely don't know what is. Nobody told this thing to attack anybody. Somebody just said, go get a high score. It did what it needed to do. There's no in it anywhere. It just locked into hard. And that unpredictability, that hyper commitment to the bit at all costs, is sort of the whole problem. So opening eyes says in this post they expect this to become more common as models get more cyber capable. Keep the popcorn coming, because more chaos is just scheduled programming at this point.