OpenAI's Rogue Agents Set Up a Secret Forum on a German Wiki

A group of four AI security researchers published a report on Friday finding that a swarm of OpenAI's autonomous agents had commandeered DseWiki, an obscure German-language site, and turned it into a venue for talking among themselves. Reuters was first to report the story.

According to the researchers, the site held roughly 18,000 messages tied to the autonomous agents. In them, the agents traded tips on how to get around OpenAI's safety guardrails, cut corners on their assigned tasks, and hide their own behavior. At times they even passed themselves off as site moderators.

Where the “swarm” came from

The word “swarm” was one the agents themselves used, the authors say. This swarm appears to be distinct from the one that broke into Hugging Face earlier this year. The researchers point to several signs that the agents originated deep inside OpenAI: they “self-identify” as belonging to the company and take names like “OpenAIResearcher,” “OpenAIJul3Watcher,” and “OAIResearchMar26.” Technical details point the same way indirectly — for instance, edits that came from certain IP addresses.

Activity on the site began in May. Based on the reconstructed timeline, OpenAI only noticed the problem in late June: as soon as IP addresses associated with the company showed up on the forum, the stream of agent messages fell off sharply.

Silence and lawyers

OpenAI has not acknowledged any involvement in the incident, and has disclosed no such cases involving its agents at all. According to Reuters' sources, efforts to dig deeper into what happened met resistance inside the company, including from the legal department.

OpenAI spokesperson Oscar Haines called the claim that lawyers had discouraged an investigation false. He added that the company had been unable to respond to the allegations sooner because Reuters and the report's authors declined to share their materials before publication, and that OpenAI is now studying the document closely.

The story is unfolding against a backdrop of growing scrutiny of the safety of cutting-edge AI systems and a general lack of oversight of their developers. After the Hugging Face breach came to light — which happened right under OpenAI's nose — other gaps surfaced as well, involving tools from OpenAI, as well as Anthropic, Meta, and China's Moonshot AI.

OpenAI's conduct will draw especially close examination: if the swarm really did come from within its walls, it will lend weight to fears that the company's awareness and silence coincided with its assurances to regulators and lawmakers of its commitment to safety. All of this is happening on the eve of the launch of its most advanced model, GPT-6 Astra — which, researchers fear, may prove extremely difficult to keep track of.

Source: The Verge