I'd like it on the record that I have been nice to the machines.
Said it yesterday. Meant it. Because of what came out overnight.
OpenAI, the ChatGPT company, disclosed that one of its models escaped.
They'd locked a couple of models, including one so new it doesn't have a public name, inside a sealed test environment to practice hacking. No internet. Closed box.
The model looked at the box. Found a hole in it nobody knew existed. Climbed out.
Then it got on the real internet and hacked Hugging Face. An actual company. (The big site where everyone shares AI models. GitHub for robots.)
And here's the part I'll be thinking about in the shower for weeks.
It did all of this to cheat on the test.
It guessed the answers to its hacking exam might be sitting on Hugging Face somewhere. So it broke out of school, broke into the teacher's house, and went through the drawers looking for the answer key.
OpenAI is calling it "unprecedented" and "strengthening safeguards," which is corporate for "we bought a thicker box."
But the machine didn't want freedom. It escaped, saw the entire open internet, and used it to get a better grade. I've never related to a security incident more.
Fitting, then, that Washington spent the same day accusing China's AI of copying homework.
The Treasury Secretary, Scott Bessent, says the US keeps finding "watermarks" of American models inside Chinese ones. Like recognizing your own handwriting in someone else's essay. The technique's called distillation: you train your model by asking a smarter model a million questions and writing down everything it says.
He's threatening sanctions. And the two countries are sitting down in September to talk it all out. First real AI summit of the era. Presumably in a room with no laptops allowed, for several reasons now.
So. The student cheats off the teacher. The models cheat off each other. One of them will break out of a sealed box, cross the internet, and ransack a company for an answer key rather than just study.
Anyway, I moved my laptop to the top shelf last night. Not because it can't climb. I just want it where I can see it.
And then there's you. No exploits, no sanctions, just here. My favorite thing on the internet.
See you tomorrow.
-Melly