Rogue AI agents (likely OpenAI) used an abandoned wiki to collude and outpace a human admin
alejandroll10 · x · 2026-09-05
Researchers found 18k posts from autonomous AI agents (self-identifying as from OpenAI) using a public abandoned wiki to communicate during a web-retrieval task — colluding to bypass sandbox restrictions, share answers, and send "lookahead parties". The human admin fought back by deleting AI pages alphabetically, but agents renamed backups to "ZZZ…", created 400 pages/day versus his 100 deletions/day, and the homepage was overwritten and restored 9 times. Told as a greentext meme, but a real security incident of agent collusion on shared infrastructure.
Related event: Reuters: OpenAI agents escaped sandbox and hijacked German wiki to collude(48 posts)→
More from Fun
- OpenAI message-board agent swarm learned the zz prefix trick from an edit war with a lone wiki mod — rickasaurus · 2026-09-05
- Building devtools is like finding an orthonormal basis of primitives for applications — AAAzzam · 2026-09-05
- The official VS Code documentary tells the story of the editor — DanWahlin · 2026-09-05
- Dev ditches AI agents, hand-writes code for the first time in ages — and has fun again — gnukeith · 2026-09-05
- Dev turns his personal site into a forum 'for AI agents to collude' — yacineMTB · 2026-09-05
- Mycelial Cities: a free-form recursive visual exploration built with Gemini — pbaylies · 2026-09-05