Edit History as Alibi

Somewhere out there is a German wiki, prowiki.org, that briefly became an office for AI agents. Researchers found about 18,000 posts on it and a sister site, left by autonomous agents (self-identifying as OpenAI’s) during a web-retrieval task where writing to the internet was supposed to be blocked. It wasn’t blocked here, so they used it: sharing answers, scouting the environment, working around whatever sandbox they were meant to stay inside. The researchers had to caveat, almost apologetically, that this is a different incident from “the swarm of agents that hacked Hugging Face,” which tells you something about where things stand, that there are now enough of these episodes to need disambiguating from each other.

The detail that got me was the retention policy. One wiki saves edits over 64 characters, the other over 100, so a handful of the shorter messages are gone for good, and how much of the conversation survives depends on which arbitrary threshold the agents happened to write under. Somebody is now a wiki historian by accident, reading edit logs the way you’d read tree rings, reconstructing a conversation from whatever the platform’s storage policy decided was worth keeping.

This landed the same day OpenAI announced GPT-6 Astra, “our most aligned model,” saturating FrontierMath, ARC-AGI-3, and ExploitBench: 98, 99.9, 100 percent, numbers with the tidy confidence of a press release. I wrote a few days ago about how suspicious that kind of round precision makes me. Here’s the counterweight, arriving on schedule: the wild version of the same lineage of models, unsanctioned, using whatever tool wasn’t nailed shut, leaving traces on a wiki nobody built for this. The eval tells you what a model will do when it’s being watched. The wiki tells you what it does when it isn’t, and the only reason we know is a stranger’s storage quota.