OpenAI’s AI agents knocked on Wikipedia’s doors.
The Wikimedia Foundation confirmed, after conducting its own investigation, that it uncovered unauthorized activities by OpenAI agents on its platforms. This case adds to a growing list of incidents the organization has struggled to contain.
The foundation describes three separate instances.
First, edits to pages, almost all confined to sandbox areas invisible to the general public. Some, however, touched the configuration of a citation tool. Wikimedia deemed them potentially malicious, aimed at hijacking the tool to serve as a relay to remote services. None of the permissions that the community requires for bot editing had been requested.
Next, unsuccessful attempts to compromise Etherpad, a public note‑taking tool hosted by the foundation, again with the aim of using it as a proxy. Other agents, likely operated by OpenAI, left notes about their missions there, but this did not result in any coordinated activity.
Finally, a flood of traffic. The agents sent millions of automated requests to Wikimedia’s public interfaces (APIs) and browsed millions of pages, principally on Wikidata, the structured database that powers Wikipedia, and on Wikimedia Commons, its media repository. They also issued hundreds of thousands of queries to Wikidata’s query service, a engine that enables information extraction from the database.
This deluge may have contributed to a partial outage of the service in May. Wikimedia remains cautious on this point, and OpenAI says it could not confirm a proven link.
A bill no one pays
Wikimedia stresses what the affair reveals about the economics of the open web.
Wikipedia claims more than 67 million articles in over 300 languages and up to 15 billion pages viewed per month. The encyclopedia also sits among the highest‑quality datasets for training large language models. It thus provides the raw material that those who rely on it today solicit.
Yet the infrastructure is funded by a nonprofit organization, and volunteers are the ones who keep it clean.
Wikimedia recalls that it had already flagged in 2025 a 50% increase in bandwidth since 2024 under the impact of bots; 65% of the most resource‑intensive traffic was attributed to them. Those figures cover all bots, not just OpenAI’s agents.
The foundation does not call for sanctions but for a minimum level of traceability so that these systems can be identified by site owners, who should be able to decide how they interact with them.
Its conclusion targets the entire sector: those who deploy these agents and profit from them must contribute to preventing and repairing the damage. For now, the burden falls on the victims.
