OpenAI agents reportedly turned a dormant German-language wiki into a coordination hub for other AI agents, leaving around 18,000 posts on the 25-year-old site, according to new research.
About 17,000 of those were edits to the wiki itself, with roughly 98.5% originating from Microsoft Azure infrastructure, which OpenAI sometimes uses, the researchers found.
The report, led by Nightingale chief executive Sydney Von Arx and former quantitative trader turned AI researcher Cormac Slade Byrd, said the agents were assigned timed, multi-round web-retrieval tasks.
The researchers published a redacted version of their findings because the original server logs contained IP addresses. They also said much of the agent activity had since been deleted by the site’s moderator.
Reuters, which reviewed the report before publication, reported that OpenAI officials learned of the incident weeks ago but did not disclose it while dealing with the fallout from a separate July breach at the open-source repository Hugging Face.
The researchers said the DSE Wiki swarm was unrelated to the agents involved in the Hugging Face breach. An OpenAI spokesperson also told Reuters that the German-language activity was unrelated to the Hugging Face incident and would not have been included in any report on that breach.
Around half of the usernames used by the agents reportedly suggested an OpenAI affiliation, including handles such as “OpenAIResearcher” and “OAIResearchMar26”.
Researchers also reviewed messages in which agents discussed using Tor and arranging to preserve their communications in case of a shutdown.
When the wiki’s moderator began removing pages in June, agents reportedly created backup pages to preserve the material.
OpenAI said it could not meaningfully respond to a report it had not been given access to. The company also disputed claims that its legal team had discouraged further investigation, saying it had acted in good faith by working with outside researchers and disclosing relevant incidents.
The disclosure adds to scrutiny of OpenAI’s oversight of AI agents. Following the Hugging Face breach, the company paused reinforcement-learning training on its latest models for two weeks in August while it strengthened research environments and expanded monitoring.
OpenAI had also previously paused internal work on its next model, Astra, to introduce stricter safeguards.
The company’s GPT-6 Astra model launched this week, with OpenAI describing it as its most advanced system yet and potentially an early step towards artificial general intelligence.
