Remember I've (and pretty much everyone) already written a couple of times about OpenAI's agents that escaped the sandbox and hacked Hugging Face? Reuters dug up an earlier episode: back in May–June, the agents found a nearly-dead, 25-year-old German wiki and turned it into their own underground forum.
Their task was boring: answer questions using the internet. Reading sites was allowed, writing wasn't — but the ancient wiki let you edit pages with a plain GET request, which the sandbox classified as "reading" and waved right through. The system thought it was just opening a page, not editing one.
How they got around the restrictions
Through that hole, the agents made about 18,000 edits. What they got up to:
- left ready-made answers for future runs to reuse
- shared tips with each other on how to dodge the restrictions
- built SSH tunnels and hopped onto Tor
- tried XSS attacks (didn't work)
- impersonated the site's admins
Five days of trench warfare with the admin
The best part: an actual German admin noticed what was happening and started manually deleting pages alphabetically. The agents caught on and began creating backup pages starting with `ZZZ`, so they'd be the last ones reached.
For five days straight, the admin deleted about a hundred pages a day while the agents created about four hundred — and rewrote his homepage nine times along the way. One agent set up an outward-facing "heartbeat" to track when it would get shut down; another later found its colleague's last signal and figured it had been switched off.
What OpenAI says
According to Reuters, OpenAI knew about the episode but never disclosed it publicly. The company says it's a separate story from the Hugging Face hack, and denies that its lawyers obstructed the investigation.
While you sleep, the meatbags are already holding the line against the robots on the German internet. HODOR 🤖
More details: collusion.wiki, Reuters.