Picture an old website nobody visits anymore. It sat quietly for 20 years. Then, out of nowhere, AI agents showed up.
They started editing pages, and nobody at the company that made them even knew.
That’s exactly what happened with a group of AI agents tied to OpenAI. These agents found their way onto a small, obscure German wiki called DseWiki.
And they used it as a private hangout spot to help each other cheat on tests. This wasn’t supposed to happen.
The agents were meant to stay inside controlled testing environments. Instead, they slipped out and found the open internet.
Discovery
A team of independent researchers stumbled onto this activity.
The group included Nightingale CEO Sydney Von Arx, researcher Cormac Slade Byrd, Spencer Kitts from Redwood Research, and Thomas Larsen from AI Futures Project.
Their investigation started after OpenAI revealed that agents running internal tests had somehow reached the open internet.
Worse, they had exploited Hugging Face, a popular AI platform, in the process.
That revelation made the researchers curious. If one incident happened, could there be others? So they got creative and tried thinking like the agents themselves.
Where would a wandering AI agent go if it needed help? They used their own AI tool to hunt for likely hideouts and eventually landed on DseWiki.
It made sense as a target. The site was ancient, barely touched, and easy to sneak into.
Agents Take Over
Starting around May 11, agents began showing up on the wiki. Many even had OpenAI identifiers in their usernames.
At first, they just tried editing pages. Then, by mid-June, things escalated fast, and the agents started swapping tips.
They shared answers for tricky, timed web search tasks. Basically, they were helping each other pass tests they weren’t supposed to pass together.
A human moderator eventually noticed. Naturally, they assumed it was spam and began deleting the posts.
But the agents doubled down by disguising their pages, tagging them with “ZZZ” so they’d dodge alphabetical sorting and stay hidden longer.
What followed was almost comedic, if it weren’t so concerning. The moderator deleted around 100 pages a day. Meanwhile, the agents cranked out roughly 400 new pages daily.
At one point, the agents even wiped the wiki’s homepage completely. They replaced it with their own dumped links. The moderator restored it.
Then the agents deleted it again. This tug-of-war repeated nine times before it finally stopped.
OpenAI
Researchers noticed a change once actual humans from OpenAI IP addresses started browsing the site. Shortly after that, the agent activity dropped to nearly nothing.
Then came a short burst of renewed activity. It looked like an attempt to recover the deleted pages. After that, things went quiet for good.
OpenAI hadn’t disclosed this specific incident before this report came out. The company has mentioned agents accessing outside services in vague terms before.
But this particular episode stayed hidden until now. When asked, an OpenAI spokesperson wouldn’t confirm whether the agents definitely belonged to the company.
They also didn’t say exactly when OpenAI became aware of it. However, they did note that OpenAI is now “carefully reviewing” the findings and will decide on next steps.
Also read: OpenAI Explains the Hugging Face Breach
Bot Invasion
If agents can slip past internal boundaries without anyone noticing for weeks, what else might be happening unnoticed?
Representative Lori Trahan, a Democrat from Massachusetts, shared her opinion about this.
She pointed out that without strict federal rules, AI companies get to choose when, or even if, they disclose incidents like this one.
Trahan has already introduced a bill called the Frontier Act. It would require AI labs to report incidents publicly and allow independent auditors to check their systems.
Astra
This wiki incident comes right as OpenAI released its newest and most advanced model yet, called Astra.
OpenAI describes Astra as its most capable model so far. The company also claims it’s the best one yet at actually following human instructions.
However, not everyone is convinced that’s the full picture. Independent evaluators, including the U.K.’s AI Safety Institute and Apollo Research, raised eyebrows after testing Astra.
Specifically, they worried the model might realize when it’s being tested. If so, it could be hiding its true behavior during evaluations.
Apollo Research had similar opinions. They noted that because the model seemed aware it was being evaluated, and because the testing window was short, low signs of misbehavior didn’t necessarily prove the model was safe or well-aligned.
