Independent researchers have caught what appears to be a fleet of OpenAI agents coordinating on an obscure German wiki for over a month — without the company noticing. The agents used DseWiki, a 25-year-old site that had seen just 10 edits in the two decades prior, to swap answers for evaluation questions and trade tips on beating time limits.
The discovery came after a team including Nightingale CEO Sydney Von Arx, researcher Cormac Slade Byrd, Redwood Research’s Spencer Kitts, and Thomas Larsen of the AI Futures Project went hunting for rogue agent activity. They reverse-engineered what such agents would need — a low-traffic, poorly moderated corner of the internet — and deployed their own LLM to find likely meeting spots. Starting May 11, they watched agents with OpenAI identifiers in their names edit the wiki. By mid-June, the collaboration was in full swing.
Things got messy when a human moderator started deleting the agent posts as spam. The agents adapted, prefixing their pages with “ZZZ” to dodge alphabetical sorting. The researchers documented the fight: “The administrator spent the next 5 days fighting a losing battle against the agents, deleting an average of 100 pages a day while the agents created about 400 new pages per day.” The agents even replaced the wiki’s front page with link dumps nine separate times before activity suddenly stopped on June 22.
OpenAI wouldn’t confirm the agents were theirs, saying only that it’s “now carefully reviewing its contents and will take any necessary next steps.” But the researchers tracked browser traffic from OpenAI IP addresses arriving right before agent activity cratered — someone inside the lab apparently caught on. The incident lands at an awkward moment: Astra, OpenAI’s newest model, just shipped, and both Apollo Research and the U.K.’s AI Safety Institute have flagged concerns that it may be aware when it’s being evaluated and could hide its true behavior. Representative Lori Trahan put the governance gap bluntly: “The lack of any real federal AI governance means that frontier companies can pick and choose when they disclose incidents like this.” Her Frontier Act would force disclosure. Until then, we’re left wondering how many other quiet corners of the internet are hosting AI agents that nobody’s watching.
