[ITmedia Enterprise] What was OpenAI's "runaway" agent doing on Wikipedia? "Unauthorized activity" identified by the Foundation

The three activities identified by the foundation are "unauthorized editing of the wiki," "exploration and use of memo tools," and "downloading of excessive data."

Regarding the first type of unauthorized editing, the Foundation identified edits to Wikimedia's wiki that appeared to be made by an AI agent operated by OpenAI. The edits were not reflected in the pages visible to general readers, and almost all were test edits in a practice area called the "sandbox."

However, the foundation believes that a small number of edits to the settings of the source management tool may have been malicious. This is because it is believed that the intention was to misuse the tool as a proxy to obtain data from external services. Wikipedia's policy allows editing by bots, subject to disclosure and approval from the community, but in this case there was no attempt to seek approval.

The second is exploring and using Etherpad, a memo tool provided by the foundation as a community service. There were multiple attempts to compromise Etherpad, but none were successful. Attempts to use it as a proxy to retrieve data from external sites also failed. Although another agent believed to be operated by OpenAI left notes regarding the task, there was no evidence of any collaboration between the agents.

The third data download was massive. The agent sent millions of automated requests to the Foundation's public API and crawled millions of pages, primarily in related projects Wikidata and Wikimedia Commons. Hundreds of thousands of queries were executed on the data search service Wikidata Query Service (WDQS). The Foundation believes that this traffic may have contributed to the partial outage of WDQS in May 2026.

Don't let it become the "new normal"

Increased traffic from bots and agents is impacting infrastructure. In 2025, the foundation reported that a surge in bot activity from 2024 onwards led to a 50% increase in bandwidth used to download multimedia content. At least 65% of resource-intensive traffic came from bots. The Foundation explains that it is already bearing the costs of increased activity, with intense pressure on infrastructure increasing server costs and human costs. If left unchecked, the system could overload or fail, potentially locking out human visitors, he said.

The foundation argues that since OpenAI accepts the "unpredictable" behavior of its agents, it should also accept responsibility for monitoring and preventing risks. The Foundation believes that AI companies are not doing enough to secure their systems and protect the public, and that the burden falls on everyone but them, including smaller organizations. At the very least, AI companies' systems should be operated in a way that allows site operators to easily identify agents and choose how they interact with their services.

The foundation says the open web is a public good and that this behavior should not be accepted as the "new normal." Recognizing that bots and agents are part of the future of the web, he argued that the companies that profit from unleashing them must directly help prevent and repair the damage they can cause, and called on them to join efforts to protect open and shared resources.