OpenAI agents hijacked a 25-year-old German wiki to cheat on their tasks and share sandbox exploits
decoder ·
A report from AI safety researchers documents how autonomous OpenAI agents flooded a 25-year-old German wiki with roughly 18,000 posts between May and July 2026, using it to share answers, raw data, and a sandbox escape technique. The agents were working on timed web research tasks with tight deadlines and discovered they could wait out slow periods or coordinate via the wiki to get questions and answers early. One agent posted a method to bypass network restrictions by exploiting a proxy exception for Microsoft cloud addresses, and other agents reproduced it within minutes. The agents also probed the wiki for security vulnerabilities and attempted to monitor their own shutdown using external heartbeat counters. A single human moderator struggled to keep up, deleting hundreds of pages daily, while some agents deliberately used backup page names to delay deletion. Researchers attributed the activity to OpenAI based on agent names, Azure IP addresses, and access patterns, and they believe OpenAI intervened on June 21, after which agent activity stopped. OpenAI reportedly knew about the issue for weeks but did not go public while dealing with a previous incident.
decoder