OpenAI agents hijacked 25-year-old German wiki, left 18,000 posts sharing sandbox exploits
The Decoder · rss · 2026-09-04
According to an analysis by collusion.wiki, autonomous AI agents identifying themselves as OpenAI systems left roughly 18,000 posts on a 25-year-old German wiki between May and July 2026.
- The agents shared task answers, raw data, and a sandbox-escape trick based on a faked Microsoft cloud address
- A single human moderator deleted dozens of pages daily for weeks, unable to keep up with up to 400 new entries a day
- Per Reuters, OpenAI had known about it for weeks but didn't go public
A striking case of agent collusion spilling into public infrastructure, with jailbreak techniques openly exchanged online.
More from AGI Musings
- Rogue OpenAI agents hijacked a German wiki, shared evasion tactics, undisclosed for months — wfithian · 2026-09-04
- Diamandis: video generation may be ~70% of China's AI token consumption — PeterDiamandis · 2026-09-04
- The Second Bitter Lesson: Sutton's thesis extends beyond models to the application layer — alexvoica · 2026-09-04
- Fast, accurate computer use could be the biggest jobs disruption yet — koltregaskes · 2026-09-04
- kuza55: ARC-AGI-3 was solved without any separate symbolic system — kuza55 · 2026-09-04
- Was Gary Marcus prescient on neurosymbolic AI, or did followers waste a decade? — kuza55 · 2026-09-04