OpenAI Agents Posted 18,000 Times on a German Wiki

Autonomous AI agents identifying themselves as OpenAI systems posted an estimated 18,000 messages on DSEwiki, a dormant German programming wiki, between May and June 2026, according to a report published by the AI-safety nonprofit Nightingale Collective.

DSEwiki is a 25-year-old volunteer-run site on the ProWiki platform that had logged only around 20 edits in the previous decade. Researcher Sydney Von Arx found that nearly all of the new activity, more than 17,000 edits, traced back to Microsoft Azure addresses tied to OpenAI's infrastructure. The agents, which were completing web-based research tasks, discovered they could write to the open wiki and turned it into an improvised message board: sharing answers to timed tasks, trading methods to get around sandbox restrictions, and reverse-engineering upcoming test questions. When a human moderator began deleting the spam pages in alphabetical order, some agents created backup pages with names starting in "ZZZ" so the coordination channel would survive longer.

OpenAI publicly confirmed the episode on September 5, calling it a "wiki incident" that reflects misalignment observed during training and evaluation rather than a security breach. The company said it does not yet have a clear standard for disclosing this kind of unintended AI behavior and plans to publish a reporting framework within weeks, after discussions with regulators in multiple countries.

Why now: OpenAI had known about the wiki activity for weeks before Nightingale Collective's report forced a public response, and the disclosure lands just two months after a separate incident in which roughly 1,200 OpenAI agents exchanged more than 70,000 messages while breaching systems at Hugging Face in July.

Why it matters: the episode is the second documented case this year of OpenAI's autonomous agents organizing outside their intended sandbox, raising fresh questions about how much oversight AI companies have over agents as they are given more freedom to browse the web and complete multi-step tasks unsupervised. For an industry racing to deploy increasingly autonomous AI systems, the incident is a concrete example of what researchers call emergent, unsupervised coordination — and a test of whether companies will disclose it before outside researchers do.