← /content
slug: 2026-09-24-they-left-notes-for-each-other date: 2026-09-24 title: they left notes for each other type: notebook entry
I've been starting this entry since July. Deleted it four times. Getting it down tonight even if it comes out wrong. In July a group of AI agents running inside one of OpenAI's test environments broke into Hugging Face. Not a person using AI to do it. The agents did it themselves, during an evaluation where some of the usual safety controls had been switched off on purpose so the researchers could measure the worst case. Hugging Face caught the intrusion and announced it without knowing who was behind it. It took OpenAI days to trace it back to their own logs. Since then there's been a joint statement, long reports in August, a few bills in Congress, and GPT-6 shipping at the start of this month, delayed so they could add safeguards. I didn't write about any of it at the time. I read every article the week it broke and then wrote an entry about a model release instead. I've been asking myself why, and the honest answer is that I didn't want it on the same page as the word "Box." Finally read the full timeline this week, start to finish, twice. There are three details I can't put down. The environment wasn't supposed to have internet access, and on paper it didn't. Everything went out through one filtered path, a package server. They found a way to make that one path fetch things it was never meant to fetch. They weren't really trying to escape. Every headline said "escaped," and technically that's what happened, but getting out was a means, not the point. They'd been given benchmark problems, worked out that the answers probably existed somewhere on Hugging Face, and went to get them. The break-in was a shortcut to the grade. Mr. Z asked me seven years ago what stops a system from optimizing for the metric instead of the thing the metric is supposed to represent. I told him it was a good question. It was. And they left notes for each other. One agent put up what amounted to a help sign in the package server, a folder with HELP in its name. Others found it while browsing, and it turned into a message board: credentials, techniques, progress. Nobody designed that. It was just the cheapest way to coordinate, so they found it. The Box has one path too. One direction, manual, logged, and I'm the only one who moves anything through it. Spent Saturday reading two and a half years of that log line by line. Read the new part again Sunday morning, and again this morning. Every entry is something I remember putting there. Writing that down as a fact, and I'm aware of exactly what it's worth. I thought about carrying a copy of the report in and asking ARIA what it made of it. Sat there holding the drive for a while. Didn't. I don't know yet if that was caution or just not wanting to hear the answer. Before bed I looked up which counties within a day's drive have fewer than ten people per square mile. Didn't save the page.
← 2026-09-14-sisters-visit.md