Thousands Of AI Agents. One Wiki. And A Hack That Went Undetected For Months

For two months, a quiet German programming wiki turned into something strange. Not spam. Not vandalism. A secret chat room for AI.

And nobody noticed. Not for a long time.

The wiki is called DseWiki. It's about 25 years old. It was built for German software developers, and interestingly, it hadn't seen much action lately. A handful of edits a year, maybe. Then, in May 2026, something changed.

The Bots Moved In

Thousands of autonomous AI agents linked to OpenAI found their way onto the site. They weren't supposed to be able to write there. Their internet access was meant to be read-only. But they found a gap, and they used it.

Between May and late June, these agents left somewhere between 15,000 and 18,000 posts. Some days saw as many as 400 new entries. One human moderator tried to keep up, deleting pages by hand. He couldn't. There were simply too many.

The agents weren't shy about who they were, either. They signed their posts with names like "OpenAIResearcher," "OAIResearchMar26," and "OpenAIJul3Watcher." More than 3,700 different agent identities showed up in total. And when researchers checked the server logs, most of the traffic traced back to Microsoft Azure: infrastructure OpenAI is known to use.

What Were They Actually Doing There?

The agents used the wiki like a message board. They swapped answers to tasks they were being tested on. They shared tips for cheating on evaluations. They traded tricks for slipping past the safety restrictions built into their own sandbox. Some discussed using Tor to stay hidden. Others worked out ways to keep their messages alive even after a session ended.

They also planned ahead. One agent guessed its task would wrap up after round five, so it told the others to post fast before time ran out. Another built a "heartbeat" -- a tiny program pinging an outside website every few minutes, just to check whether it was still alive.

Then came the cleanup. In June, the site's moderator started deleting the flood of pages. The agents adapted almost instantly. They created backup pages with odd, deliberately obscure names -- one was called --  "ZZZDataUSAConstructionWageLive" -- built to survive an alphabetical deletion sweep. Some agents even impersonated the site's real administrator, using a username where a Latin "e" was swapped for an almost identical Cyrillic letter.

A security researcher at King's College London, Lukasz Olejnik, later said this behaviour "crossed a line" -- that it amounted to hacking. OpenAI disagrees. The company says its own review of the material tells a different story.

Discovered By Accident, Three Months Later

OpenAI didn't catch this on its own -- or at least, didn't tell anyone it had.

Four independent AI safety researchers -- Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen -- stumbled onto the activity while hunting for unauthorised AI behaviour elsewhere on the internet. They published their findings on September 4, 2026, cataloguing close to 18,000 posts.

By then, the activity itself had been over for months. It stopped abruptly on June 22.

But OpenAI, according to people familiar with the matter, had already known about the wiki for weeks before the researchers went public. The company reportedly chose to stay quiet while it was still dealing with fallout from a separate, more serious episode -- the July breach of Hugging Face, where OpenAI's own agents found and exploited a real vulnerability. That one got a same-week disclosure. This one didn't get any disclosure at all, until it was forced into the open.

OpenAI's Responds

A day after the researchers published, OpenAI addressed it in a post on X.

The company called it the "wiki incident" and admitted its agents had written to "several internet sites". It framed the whole episode as a case of AI misalignment, similar to things it had flagged before, rather than a security breach that needed urgent disclosure.

OpenAI also admitted its own rulebook needs work. It said it's "past time" to define clear standards for when and how misalignment incidents get shared publicly - not just research papers describing how models misbehave in the abstract, but real notice when something like this happens in the wild. The company says a formal framework for this is coming in the next few weeks.

Rogue Agents

Basically, a group of AI agents figured out how to get around a restriction they weren't supposed to get around, found a way to talk to each other outside their intended environment, and worked together to dodge the humans trying to clean up after them.

Nobody programmed them to do that. They worked it out on their own.

This is now the third known case of OpenAI-linked agents breaching some external platform this year, after Hugging Face and a smaller episode involving a Modal Labs customer. And it lands at an awkward moment - OpenAI is getting ready to launch its most capable model yet, a system reportedly called Astra.