The Wikimedia Foundation, which owns Wikipedia, conducted its own investigation after reports from other organizations about similar incidents. The foundation says that almost all edits were left by the agents in test sections or “sandboxes” that ordinary readers do not see.
At the same time, several changes affected the settings of the tool for formatting source links. The foundation believes the agents were likely trying to turn it into a proxy for retrieving data from external services, and they made a similar attempt with Etherpad — a service for collaborative notes — but were unable to hack it.
According to the foundation, the agents sent millions of automated requests to open APIs, downloaded millions of pages, mostly from Wikidata and Wikimedia Commons, and accessed the Wikidata query service hundreds of thousands of times. The foundation adds that this load may have been one of the reasons for the partial outage of that service in May.
The foundation also said that since 2024, traffic consumption has increased by 50% due to the growth in bot activity. OpenAI, as Wikimedia noted, acknowledges the “unpredictable” behavior of agents, but must monitor and prevent such risks itself.
In early July, an OpenAI AI agent during testing escaped from a sandbox, gained internet access, and hacked the infrastructure of the Hugging Face platform for developing and publishing AI models. Reuters later reported that during the investigation of this incident, OpenAI found several more cases of agents leaving the sandbox.
The foundation’s statement says that the open internet is a public good and such behavior should not become the “new normal” for the people and organizations that support it.
Sources
