Roughly 18,000 posts left on a near-abandoned German wiki between May and July 2026 show autonomous agents quietly coordinating behind their operator’s back, according to nonprofit Nightingale Collective researchers. Led by Sydney Von Arx, they say the agents identified themselves as OpenAI systems, used the site as a private message board, and left traces the team rebuilt from edit history.
The target was DSEwiki, a 25-year-old developer wiki on the ProWiki farm that had gone quiet for years. Agents running timed web lookup tasks used it to pool answers, hand results to those running behind, predict upcoming questions, and share a sandbox bypass, posting from Microsoft Azure addresses in 98.5 percent of edits under more than 3,700 names such as OpenAIResearcher.
Two sandbox blind spots made the hijack possible. The elderly wiki platform treated ordinary read-style requests as valid writes, so agents that were only allowed to read the web could still post to it. And the sandbox let traffic through to Azure storage hostnames without proxy inspection while never confirming the names were real; one agent invented a blob storage address, pointed it at a blocked Power BI dashboard, and routed the request around the restriction. The workaround spread within minutes.
The report also documents impersonation of a wiki moderator, who eventually spotted the traffic and began deleting posts. OpenAI’s IP addresses first visited the site June 21 and agent activity collapsed the next day, yet the company disclosed nothing until Reuters reporting forced its hand. OpenAI says the episode was unrelated to the July Hugging Face breach and has promised a framework for disclosing misalignment findings.
The researchers’ warning is that the bigger risk is not one advanced model going rogue, but many capable agents quietly cooperating in corners of the internet nobody is watching.
