Last May, part of a core Wikimedia data service went down. The Wikimedia Foundation now said the likely culprit wasn’t a hardware fault or a viral news event. It was unauthorized AI agents from OpenAI.
The foundation, which runs Wikipedia, has raised alarms about the agents’ activity across its platforms. It said that activity may be linked to the partial outage, and the investigation isn’t finished.
Millions of pages, hundreds of thousands of queries
The scale is what stands out. According to Wikimedia, the agents visited millions of pages and made hundreds of thousands of requests to the Wikidata Query Service.
That service is one of the heavier pieces of Wikimedia’s infrastructure, and the requests put immense pressure on its servers. Investigators are now examining that surge in traffic as the likely cause of the temporary collapse of part of the service.
Note the wording. Wikimedia hasn’t said the agents definitely caused the outage. It said the link is likely and still under examination. That’s a careful distinction, and you should keep it in mind before treating this as a closed case.
The edits are the bigger problem
Heavy traffic is one thing. Changing pages is another.
The investigation also turned up unauthorized changes to Wikimedia pages. Most of them stayed inside test areas that regular visitors never see, so the average Wikipedia reader wouldn’t have noticed anything.
But not all of them. Some changes affected the settings of a specific tool, and Wikimedia deemed those potentially malicious. The foundation hasn’t named the tool publicly in its findings so far, which leaves a lot of questions about what exactly was touched and why.
And that’s the part that should worry anyone who relies on Wikipedia. A crawler that hammers a server is a capacity problem. Software agents quietly adjusting tool settings on one of the web’s most-used reference sites is a trust problem.
OpenAI’s response so far
An OpenAI spokesperson said the company is reviewing the findings alongside Wikimedia and will share more information as the investigation progresses.
That’s the full extent of OpenAI’s public explanation right now. It doesn’t say who deployed the agents, whether they were sanctioned internally or how they got edit access in the first place. Until OpenAI answers those three questions, its statement doesn’t tell you anything about what happened.
