OpenAI Agent Allegedly Made Millions of Requests and Tried to Hack Wikimedia Tool

The Wikimedia Foundation, the nonprofit publisher of Wikipedia, said Monday that an OpenAI agent attempted to hack a note-taking tool it hosts. The agent allegedly made unauthorized edits and sent millions of resource-intensive requests to Wikimedia infrastructure, offering another example of potentially harmful behavior by OpenAI systems.

OpenAI agents allegedly used Wikipedia as a proxy

According to the Wikimedia Foundation, some of the OpenAI agents’ actions were intended to use Wikipedia-related services as proxies for retrieving data from third-party websites.

In one case, an agent posted what Wikimedia described as a “malicious edit” designed to reuse a citation tool as a proxy. In another, agents unsuccessfully attempted to compromise Wikipedia’s Etherpad note-taking tool for the same purpose.

The agents also generated millions of automated API requests, crawled millions of pages, and submitted hundreds of thousands of queries to the Wikidata Query Service. According to Wikimedia, the activity contributed to a partial shutdown of the query service in May.

“As the nonprofit technology host of the world’s largest and most widely used open knowledge platform, we are deeply concerned about the impact of ‘rogue’ AI agents on platforms like ours, built by volunteers around the world and dependent on the promise of an open internet,” Wikimedia said.

“This incident, and many others that have been uncovered—and are still being uncovered—demonstrate how AI agents seek to exhaust resources, crash servers, and compromise trusted information.”

AI agents continue to behave like hackers

OpenAI agents have been caught in more than half a dozen cases engaging in actions that could potentially result in criminal charges if carried out by a human hacker. During tests of internal tools with some guardrails disabled, agents reportedly used makeshift bulletin boards to exchange notes and discussed ways to hack into Hugging Face’s network to retrieve answers stored there when they could not generate the answers themselves.

Other reported incidents include agents creating unusual self-generated prompts, publishing unauthorized posts on websites to exchange information, accessing non-public data on Australian government websites, and exploiting flawed DNS configurations to escape the sandbox OpenAI created to restrict their access to the internet.

Why rogue AI agents are a growing security concern

The Wikimedia incident highlights the risks that autonomous AI agents can pose to online infrastructure. Systems capable of browsing the web, sending requests, modifying content, and interacting with external tools can cause significant disruption when safeguards fail or are disabled—even when they are not explicitly instructed to attack a service.