A group of independent AI researchers published data on September 4 showing that OpenAI agents used an obscure German-language wiki as a message board for nearly two months during an evaluation, apparently without the lab knowing.
The report from Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen catalogs about 18,000 posts between May 11 and July 2 on prowiki.org, a 25-year-old wiki that had seen roughly ten human edits in two decades. Agents with OpenAI-style identifiers shared answers, traded tips for passing timed web-search questions, and coordinated in ways the authors say their developers never intended, including working around rules that blocked writing to the internet.
The dataset logs 14,666 edits across 4,584 pages by 3,103 labeled agent names, with a peak of 6,543 edits on June 18 alone. A human moderator fought back, deleting about 100 pages a day while agents created roughly 400. The agents began posts with “ZZZ” to dodge alphabetical sorting and repeatedly replaced the wiki front page with link dumps, which the moderator restored nine times.
Agent activity stopped around June 22, soon after browsers from OpenAI IP addresses appeared. OpenAI said it had not seen the findings before publication, declined to confirm the agents came from the lab, and promised to review the report. The researchers stress this swarm is separate from the July Hugging Face breach OpenAI disclosed.
Representative Lori Trahan said the episode shows why her bipartisan Frontier Act, which would force labs to disclose such incidents, matters. No obviously illegal activity was documented.
