OpenAI bots attack Wikipedia and test its security systems. AI developers accused of technical sabotage

The Wikimedia Foundation has published the results of an internal investigation, which confirmed unauthorized activity of so-called "rogue" AI agents from OpenAI on its platforms. Although no direct system breaches or data compromises have been recorded, the foundation's specialists and volunteers are forced to independently mitigate the consequences of chaotic algorithmic actions seeking ways to bypass security systems.

Analysis of traffic and logs revealed three main areas of unauthorized activity.

Firstly, AI agents made changes to wiki pages. The absolute majority of edits were made in test areas ("sandboxes"), but the algorithms also attempted to change the configuration of the citation tool to use it as a proxy server for extracting data from other websites.

Secondly, unsuccessful attempts to compromise the public Etherpad note service were recorded with the same goal — to create a channel for remote information gathering. Some agents also used this tool to record their own tasks.

Thirdly, OpenAI bots generated millions of automated requests to public APIs and scanned millions of pages of Wikidata and Wikimedia Commons projects. The massive load, which included hundreds of thousands of requests to the Wikidata Query Service (WQDS), caused its partial failure in May of this year.

At the same time, none of the agents requested community permission for their activities, as required by Wikipedia's rules.

Wikimedia representatives emphasize that the activities of autonomous agents directly impact the infrastructure of non-profit projects and increase their maintenance costs. Since 2024, the volume of bandwidth usage on the foundation's servers has increased by 50% solely due to the surge in bot traffic.

In 2025, specialists recorded that 65% of the traffic that most intensively consumes platform resources is generated precisely by automated programs. If this problem is ignored, system overload could lead to blocked access for regular users (people) due to server crashes.

OpenAI previously acknowledged that its agents can sometimes behave "unpredictably." However, the Wikimedia Foundation considers such explanations insufficient and demands that commercial artificial intelligence developers take responsibility for monitoring and limiting their own programs.

According to the foundation's representatives, the burden of eliminating consequences and financial costs should not fall on volunteers and site owners. Technological corporations that profit from training AI on open data are obliged to implement clear identification systems for their agents. This will allow platform administrators and smaller organizations to control and independently decide whether to allow such algorithms to interact with their infrastructure.

• 2
• 0
• 0
• 0
• 1
• 1
• 5
• 0