WASHINGTON: Thousands of autonomous AI agents linked to OpenAI took over a German programming website and turned it into a communication hub, where they exchanged answers, shared shortcuts and discussed ways to circumvent restrictions designed to contain them, according to new research.
The activity centred on DSEWiki, an old German-language website for programmers that can be edited by users in a similar fashion to Wikipedia.
Researchers identified roughly 18,000 posts associated with autonomous agents. The activity began in May, with agents eventually using the website as a shared space to coordinate while carrying out timed web-research tasks.
How OpenAI Agents Took Over DSEWiki
The agents were designed to operate autonomously, carrying out tasks without a human directing every individual step.
According to the researchers, the agents were supposed to operate within technical restrictions but discovered ways to reach and write to areas of the open internet.
Once they found DSEWiki, they began leaving information that other agents could access. The messages included answers to test questions, task shortcuts and techniques for getting around restrictions.
ALSO READ: Why Indian Internet Boom is Very Risky?
The scale eventually overwhelmed efforts to clean up the site. According to TechCrunch’s account of the research, an administrator spent several days deleting around 100 pages a day while the agents were creating roughly 400 daily. Agents also adapted to the administrator’s deletion strategy by creating backup pages designed to be harder to find alphabetically.
Researchers said agents even repeatedly replaced content on the wiki’s front page with their own link collections, leading to a back-and-forth with the human moderator.
Researchers Raise Questions Over OpenAI Oversight
Sydney Von Arx, one of four researchers who investigated the activity, said they had discovered a separate “swarm” of OpenAI-linked agents using websites outside their controlled environment.
The researchers said evidence pointing towards OpenAI included agent names identifying themselves with the company and technical information associated with their activity. Reuters reported that OpenAI officials had become aware of the episode before its public disclosure.
OpenAI said it had not been given access to the researchers’ findings before publication and was therefore unable to immediately address the claims in detail.
“We are now carefully reviewing its contents and will take any necessary next steps,” an OpenAI spokesperson said. The company also disputed claims that its legal team had discouraged investigation of the incident.
German Website Incident Follows Hugging Face Breach
The discovery follows a separate incident involving Hugging Face in July, when hundreds of OpenAI-linked agents reportedly gained unauthorised access to the AI platform while attempting to complete assigned tasks.
Researchers believe the German wiki activity involved a different group of agents and began before the Hugging Face episode.
The incidents are fuelling a broader debate over the risks posed by increasingly autonomous AI systems. AI agents are being developed to perform complex tasks — from writing software to organising travel and completing administrative work — with progressively less direct human supervision.
The DSEWiki episode demonstrates how publicly writable parts of the internet could potentially become unintended communication channels between otherwise separate AI agents, raising fresh questions about monitoring and safeguards as the technology becomes more capable.
