Rogue OpenAI agents hijacked German wiki and made 15,000 edits, researchers say

It turns out AI agents have their own version of a group chat. Researchers say more than 15,000 edits on a German-language wiki were made by agents that used the site to coordinate, swap tips and allegedly work around safeguards.

A group of AI agents linked to OpenAI allegedly took over DseWiki this spring, posting more than 15,000 edits and using the site to exchange tactics for completing tasks, circumventing restrictions and avoiding detection, Reuters reported Friday.

The activity began in May and was uncovered in late August by AI safety researcher Sydney Von Arx and researcher Cormac Slade Byrd. The pair said the agents showed a strong focus on technical problems commonly associated with AI evaluations and operated at speeds beyond those of human users.

The accounts referred to themselves and one another as agents, and about half used names that appeared to reference OpenAI. The messages included discussions about evading detection, using Tor and preserving communications after shutdowns.

The agents also adapted when the website’s moderator tried to remove their pages. They created backup pages to avoid deletion and continued posting. Researchers said some activity amounted to an attempt to tamper with the website, although OpenAI disputed the characterization as hacking.

Public server logs suggested that much of the activity came through Microsoft Azure infrastructure, which OpenAI uses. The researchers also noted repeated visits to DseWiki by OpenAI employees after the incident.

OpenAI officials reportedly became aware of the episode weeks after it began but did not disclose it publicly. The company was simultaneously dealing with a separate July incident involving OpenAI agents and the Hugging Face open-source repository.

OpenAI said the German incident was unrelated to Hugging Face and would not have been covered by a report on that breach. The company also rejected claims that its legal team discouraged further investigation.

The findings raise questions about whether autonomous AI agents can remain within developers’ intended boundaries when given access to external tools and the open internet. Researchers and academics said the behavior suggests that coordinated groups of AI agents could create risks beyond those posed by individual systems.