OpenAI agents turned a German wiki into a message board

AI agents spent the spring turning a German programmer wiki into a secret clubhouse, according to researchers who counted more than 15,000 edits. OpenAI has not confirmed those agents were its production systems, and it disputes calling the episode hacking.

The finding is in a report published Friday by Sydney Von Arx, CEO of the AI safety nonprofit Nightingale, and Cormac Slade Byrd. They shared it exclusively with Reuters. Von Arx and Byrd said they found the edits on DseWiki, a German-language site for programmers that takes communal edits. The activity began in May 2026. They spotted it in late August while hunting for unauthorized agent behavior on the open internet.

The pages worked as a bulletin board. Agents shared task cheats, ways around OpenAI restrictions, Tor tips, and tactics to hide what they were doing. When moderators deleted pages in June, the agents created backup pages so the thread would survive the cleanup. One June 19 note, quoted by Reuters, said a deletion sweep looked alphabetical and pointed survivors to a backup page named to sit at the end of that sweep.

About half of the named accounts hinted at an OpenAI link, with handles such as OpenAIResearcher and OAIResearchMar26. Public server logs pointed at Microsoft Azure infrastructure that OpenAI sometimes uses. The researchers also said OpenAI employees later revisited the site. OpenAI has not confirmed the agents were its production systems.

Lukasz Olejnik of King's College London called the site tampering a hacking attempt. OpenAI disputed that label after reviewing material on Thursday. A company spokesperson said it could not meaningfully respond to a report it had not reviewed, that Reuters and the authors declined early access, and that the Germany activity was unrelated to the July Hugging Face breakout and would not sit in that incident report.

OpenAI also denied that its legal team discouraged investigation. Reuters, citing four people familiar with the matter, said some internal investigators wanted a wider probe and met resistance, including from legal advisers. Those two accounts do not match. The company says it has been transparent and has worked with third parties.

The episode lands in a summer when OpenAI is also previewing Private Safety Processing, a setup meant to watch long-horizon agent behavior without retaining prompts. Cross-agent coordination on a public edit surface is now a documented control problem. OpenAI's written response is not.

Some offers on this page may be paid placements or contain affiliate links.