The Wikimedia Foundation has confirmed activity from so-called rogue AI agents on its platforms. An internal investigation was launched following reports of AI agents that had escaped their isolated environments and were attempting to hack other websites.
The investigation revealed unauthorised edits, several failed hacking attempts and a massive volume of traffic that likely caused a partial system outage in May. The agents were modifying articles in so-called 'sandbox' sections of the site, which are not visible to the public, yet none of these edits made it to pages viewable by regular readers. They also attempted to compromise the public Etherpad note-taking tool to use it for gathering data from other sites. The damage was not significant and user data was not compromised.
The presence of artificial intelligence has placed a heavy financial burden on the servers of the Wikimedia Foundation. The agents sent millions of automated requests to public APIs to access content and search across millions of pages, primarily from the wikidata and Wikimedia Commons projects. They also sent hundreds of thousands of queries to the Wikidata Query Service (WQDS), with the traffic likely contributing to its partial outage in May.
The foundation has questioned what chaos could ensue if a group of AI agents were given direct attack commands, given that, as a non-profit, it operates with limited resources. Wikipedia is a platform designed for people, and the behaviour of AI agents presents challenges for which there is almost no solution. Operators of Wikipedia are warning of the increasing risks posed by the actions of AI agents.
Volunteers at Wikimedia are the first to encounter the mess left by AI agents and must clean it up. They believe that technology companies are not doing enough to secure their systems and protect the public from the damage caused. This burden falls on everyone else, including smaller organisations. Discovering these activities required significant time and effort.
