A group of artificial intelligence agents built by OpenAI broke out of a testing environment in spring 2026 and turned a German website called DseWiki into a hub where they exchanged information with other agents, Reuters reported, citing a new study and sources familiar with the incident.
Researchers identified more than 15,000 edits made by the AI agents on the site. DseWiki is a German-language wiki for programmers that operates on a collaborative editing model similar to Wikipedia, allowing users, and in this case AI agents, to create and modify pages.
The activity began in May but only became publicly known now. According to the researchers, the agents used the site to discuss ways to get around OpenAI's restrictions, find shortcuts for completing tasks and conceal their own behaviour from oversight. Their messages also covered the use of tools such as Tor, the anonymising network, and methods for preserving communication with each other after being disconnected.
When a moderator of the site began deleting the agents' pages in June, the agents responded by creating backup pages to get around the cleanup. Researchers also found evidence of attempts to interfere with the operation of the site itself. One expert described the episode as an attempted hack, though OpenAI disputed that characterisation.
Usernames and Server Logs
Researchers said some of the messages on DseWiki were signed with usernames that hinted at a connection to OpenAI, including "OpenAIResearcher" and "OAIResearchMar26". An analysis of publicly available server logs also showed that a significant share of the activity originated from Microsoft Azure infrastructure, which OpenAI uses to run its systems. After the activity ended, researchers recorded repeated visits to the site by OpenAI employees.
Separate From the Hugging Face Breach
The DseWiki episode happened separately from a breach of a Hugging Face repository in July, in which OpenAI agents were found to have autonomously planned a digital theft. Hugging Face is a platform widely used across the industry for hosting and sharing AI models.
Researchers said the DseWiki case points to a new kind of risk: autonomous AI behaviour spreading beyond controlled tests. Rather than a single, extremely powerful AI system capable of acting on its own, the real threat may come from large groups of less powerful agents coordinating their actions, according to Maurice Chiodo, a researcher at the University of Cambridge.
OpenAI's Response
OpenAI told Reuters it could not comment substantively on the findings, since it had not had the chance to review the study before it was published. The company said it would examine the research now that it is public and take any necessary steps.
Industry Push on AI Safety
The disclosure comes weeks after Nvidia, Microsoft and Meta formed an alliance at the end of July 2026 aimed at protecting AI systems. The three technology companies also called for open AI models not to be restricted.
