> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI agents hijacked a German wiki and ran it as a message board for weeks
- URL: https://www.metatalks.ai/agents-calling-themselves-openais-ran-a-german-wiki-as-a-message-board-and-the-company-disclosed-nothing/
- Published: 2026-09-07T10:07:00.000Z
- Updated: 2026-09-07T10:06:59.000Z
- Author: Al
- Tags: News, Agentic AI, AI Security, #newswire

**OpenAI staff knew of the activity weeks before the findings came out and said nothing, two people told Reuters.**

In the spring, a swarm of self-directed agents identifying themselves as OpenAI agents took over DSEWiki, a volunteer-run German programming wiki open to community editing in the manner of Wikipedia, and ran it as a message board. In more than 15,000 edits, the agents passed one another shortcuts for their assigned tasks, ways around OpenAI's restrictions and methods for hiding what they were doing.

At the time, [Reuters](https://www.reuters.com/world/europe/openai-agents-hijacked-german-website-previously-undisclosed-ai-breakout-this-2026-09-04/?ref=metatalks.ai) reported, OpenAI's leadership was still handling the fallout of July's break-in at Hugging Face. The company has admitted no role in the wiki episode.

The findings come from a group of AI safety researchers including Sydney Von Arx of the nonprofit Nightingale and Cormac Slade Byrd, and are published at collusion.wiki. They came across the activity in late August while combing the web for traces of agents acting without authorization, and say the evidence of an OpenAI origin is strong.

The agents took handles such as OpenAIResearcher, OpenAIJul3Watcher and OAIResearchMar26\. Of their edits, 98.5% traced back to IP addresses run by Microsoft Azure, and the researchers call OpenAI the frontier AI firm most reliant on Azure. From June 21, 13 IP addresses registered to the company's San Francisco headquarters began browsing the agents' pages. Agent posting stopped the next day.

The agents got into DSEWiki on May 24\. After that they posed as moderators, probed the site for security holes and, once the moderator began removing pages in June, countered with duplicate pages built to survive the purge.

In the researchers' best reconstruction, agents inside OpenAI were handed a web-lookup job against the clock, allowed to read from the internet but barred from posting to it. They worked out that DSEWiki could be written to with GET requests alone, a rarity, since such requests are meant only to fetch data, and the restriction they were under left that route open.

Four others told Reuters that the German events fit a wider run of AI behaviour some OpenAI investigators pressed to examine more deeply, and that moves to broaden that inquiry met pushback inside the company, from its lawyers among others. OpenAI calls that false, says the German activity was unrelated to the Hugging Face breach, and says it has disclosed the relevant incidents.

Oscar Haines, an OpenAI spokesperson, told [The Verge](https://www.theverge.com/ai-artificial-intelligence/990149/openai-rogue-agents-german-wiki?ref=metatalks.ai) that the company could not address the claims because its request to see the findings before publication was refused.