Wikimedia flags unauthorized OpenAI-linked agents targeting its infrastructure

Wikimedia Foundation detected unauthorized automated activity carried out by artificial intelligence agents operating from an environment tied to OpenAI, the nonprofit revealed in an investigative report released Monday.

The rogue operations included test edits across wiki sites, probes targeting internal tools, and millions of automated requests that flooded Wikimedia servers, though the foundation stated the incidents did not breach its core systems or leak user data.

Unusual traffic spikes from these automated agents may have contributed to a partial outage that hit the Wikidata Query Service last May, the foundation noted, while stressing that investigators found no evidence anyone compromised its infrastructure or used it to coordinate agent swarms.

Most of the automated edits occurred in test sandboxes and subpages shielded from general public view. Investigators, however, also tracked modifications to citation tool settings—an intervention Wikimedia suspects was an attempt to turn the tool into a proxy relay to extract data from external services.

The unauthorized agents also mounted unsuccessful attempts to access Etherpad, a community note-taking tool that Wikimedia hosts for collaborative work. None of the software involved secured the required permissions or community approvals mandated by Wikipedia policies governing automated bot accounts.

Tech publication The Record reported that Wikimedia’s findings draw a hard line between abusive resource consumption and a network compromise, confirming that monitors recorded no data leaks.

The incident underscores the crushing burden that automated scrapers deployed by commercial AI firms continue to offload onto open-web infrastructure. According to the report, the agents fired off millions of automated requests, crawled millions of pages across Wikidata and Wikimedia Commons, and ran hundreds of thousands of high-load computational queries against Wikimedia databases.

Data consumption across Wikimedia servers jumped nearly 50% this year alone, driven by an aggressive surge in automated bot traffic since 2024. These automated tools now account for roughly 65% of the foundation’s most resource-intensive traffic, saddling the nonprofit with soaring operational costs and threatening to crowd out human researchers and everyday readers during peak traffic jams.

OpenAI told The Verge that it appreciated the information Wikimedia shared and is collaborating with the foundation to analyze the activity.

The San Francisco-based tech firm did not independently confirm whether its autonomous agents triggered the May Wikidata crash, a point Wikimedia itself left open, noting only that the machine-driven deluge likely functioned as a contributing stressor rather than the sole culprit.