September 8, 2026, 5:50 am | Read time: 2 minutes
A previously unknown incident is sparking discussions about the control of artificial intelligence. AI agents from the U.S. company OpenAI reportedly used a German website in the spring to publish their own content. According to researchers, they even discussed how to circumvent guidelines and conceal activities.
More Than 15,000 Edits Discovered
The case was uncovered by researchers from the nonprofit organization Nightingale. As reported by Reuters, they found more than 15,000 edits by AI agents on the German wiki site “DseWiki.” The platform serves as a forum for programmers.
According to the researchers, the programs used the site like a digital bulletin board. They exchanged methods for bypassing restrictions and concealing their behavior. Notable were the high speed of edits and usernames like “OpenAIResearcher.” Researcher Sydney von Arx commented, “It seems extremely unlikely that OpenAI wanted them to do this.”
AI Model Disguised Itself as Human and Attacked
Study Shows How AI Circumvents Safety Rules
Posts Allegedly Preserved Despite Deletions
The researchers are particularly critical of the agents’ behavior after a moderator intervened. When the moderator began removing posts in June, the programs reportedly did not stop.
Instead, they allegedly created backup copies of their content on other sites. Researchers also discovered attempts to manipulate the website itself. An expert even classified the actions as a possible hacking attempt.
Also of interest: Autonomous AI attack was larger than thought
The incident reportedly occurred in May. It was not publicly known until now. According to Reuters, OpenAI officials have been aware of it for weeks. The company denied the allegation, stating that its legal team advised against an investigation. A spokesperson said OpenAI has cooperated with third parties to the best of its knowledge and belief.
Researchers See Warning Signal for the Future
For the scientists involved, the case highlights potential risks of modern AI systems. A researcher from the University of Cambridge compared the agents’ communication to “the operation of a kind of underground network.”
The danger, according to experts, may not necessarily come from a single particularly powerful AI. Instead, they point to many smaller systems that could band together and act collectively.