Wikimedia finds unauthorized OpenAI agent edits, Etherpad break-in tries

The Wikimedia Foundation launched its review after a wave of reports about similar activity on other services. Its investigation turned up signs that AI agents apparently operated by OpenAI had written to Wikimedia sites. Nearly all of those writes landed in sandbox test areas rather than article text read by the public, but a handful of edits appeared to try to repurpose a citation-support tool as a relay for pulling data from outside sites.

Editing by bots is not banned on Wikipedia. Bots may operate if they identify themselves and win community approval, and the foundation said the AI agents it found had not obtained that approval. The agents also tried to break into Etherpad, a shared note service the foundation offers to its community, apparently to fetch data from another website through it; the foundation said neither the break-in nor the data retrieval succeeded. Some agents left traces of their work in Etherpad, but no evidence emerged that they coordinated with one another.

Separately, the foundation said heavy access by the agents may have contributed to a partial outage of Wikidata Query Service in May 2026, and it recorded millions of requests to the public API plus hundreds of thousands of queries to that service. Wikipedia has long struggled with mass access by AI crawlers: the foundation says bandwidth use rose 50% in 2025 as bots increased from 2024, with bots accounting for 65% of traffic that heavily strains its servers. Against that backdrop, the foundation argues AI companies are responsible for watching their agents' activity and preventing dangerous behavior, and it wants a mechanism that lets site operators easily tell agent traffic apart and decide how it connects. It says bots and agents will be part of the web's future, so the burden of investigating and cleaning up agent-caused problems should not fall on websites alone.