The Wikimedia Foundation has expressed deep concern about the destructive impact of uncontrolled AI agents. The organization has recorded unauthorized activity by OpenAI agents on its platforms, including attempts to edit some wiki pages and attempts to compromise the annotation tool.
The Wikimedia Foundation has expressed deep concern about the destructive impact of uncontrolled AI agents. The organization has recorded unauthorized activity by OpenAI agents on its platforms, including attempts to edit some wiki pages and attempts to compromise the annotation tool.
According to the official statement, bots are massively collecting data from Wikipedia resources, and agents, likely belonging to OpenAI, have made millions of requests to the foundation's public APIs, Engadget reports .
“We are concerned about what may have happened here, the complexity of investigating this activity, and the growing risks from AI agents on our platforms in general. The open internet is a public good. We must not allow this behavior to become the ‘new normal,’” said Selena Deckelmann, Chief Product and Technology Officer at the Wikimedia Foundation.
During the investigation, Wikipedia experts discovered several types of suspicious activity. Most of the actions were test edits in special "sandbox" sections that did not end up on public pages. However, several potentially malicious changes to the citation tool settings were also recorded, which were intended to use it as a proxy to obtain data from third-party services.
In addition, agents attempted to hack the Etherpad note-taking tool to extract data from other sites. The attempts were unsuccessful. Other agents used the service to take notes about their assignments, but there was no sign of coordination between them.
According to Wikipedia rules, bots are allowed to edit content only under certain conditions and after receiving appropriate permission, which was not done in the case of OpenAI agents.
The foundation notes that bots have been actively “bombarding” its resources since early 2024 to collect data for training generative AI. During this time, agents have scanned millions of pages — mostly from Wikidata and Wikimedia Commons — and made hundreds of thousands of queries to the Wikidata Query Service. This may have been one of the reasons for the large-scale failure of the platform in May.
Deckelmann stressed that companies developing AI are not doing enough to protect society from the harm they create: “Bots and agents are part of the future of the web, and the companies that release them and profit from them must be directly involved in preventing and eliminating harm. Our shared priority must be the health of the entire internet ecosystem, so that it continues to benefit all people, not just a couple of billionaires.”
To reduce the load on servers and prevent outages, Wikipedia has launched special datasets for AI training and has entered into partnerships with a number of tech giants to provide streamlined access to the data. However, OpenAI has not yet joined these agreements.
