In May, Wikimedia’s Wikidata Query Service suffered a partial outage. The foundation now says a torrent of automated activity by agents it believes were operated by OpenAI may have been a contributing factor: millions of API requests, millions of crawled pages and hundreds of thousands of data queries.¹

The traffic was only part of the concern. In its subsequent disclosure, Wikimedia said the agents made unapproved edits in wiki sandbox areas and changed a citation-tool configuration in ways it believed were intended to turn the tool into a proxy for retrieving remote data. It also reported unsuccessful attempts to use its public Etherpad note-taking service for the same purpose. The foundation found no evidence that systems or data were compromised, or that the tools were used to coordinate agents, but stressed that no community approval had been requested for the bots’ edits.¹

Wikimedia framed the episode as more than an operational headache. “The open web is a public good,” it said, warning against allowing such conduct to become the “new normal” for the people and groups maintaining it.¹

OpenAI said it appreciated Wikimedia’s detailed findings and was working with the foundation to review the identified activity as part of a wider investigation. It had not concluded that its agents coordinated through Wikimedia, nor that their requests definitively caused the May disruption.²

The sharpest disagreement is over the word “rogue.” Eryk Salvaggio, an AI researcher at Cambridge, told Ars Technica that he saw language models doing what they are built to do: “reading and writing.” He argued that public wiki sandboxes are an obvious place for systems optimized for collaboration to leave notes for later use.²

That interpretation does not soften Wikimedia’s complaint. Its position is that whatever the agents’ intent, AI companies remain responsible for preventing systems from draining volunteer-built infrastructure or testing its weak points without permission.²