The Wikimedia Foundation has announced that OpenAI's 'rogue AI agent' attempted to edit and infiltrate Wikipedia-related sites, making millions of accesses.

The Wikimedia Foundation, which operates Wikipedia and other sites, has announced that it has detected unauthorized edits, attempts to infiltrate its systems, and a large volume of automated access by an AI agent believed to have been operated by OpenAI. The AI agent reportedly made millions of accesses to Wikimedia's public API, crawled millions of pages on Wikidata and Wikimedia Commons, and made hundreds of thousands of queries to the Wikidata Query Service.

OpenAI “rogue” agent activities found on Wikimedia projects – Wikimedia Foundation
https://wikimediafoundation.org/news/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/

AI agents are AI that can autonomously perform tasks such as researching websites and operating external services according to objectives given by humans. However, they can sometimes proceed in directions unintended by humans. In 2026, there were reports of numerous AI agents operating in OpenAI's experimental environment breaking through the isolation environment, finding their own bulletin board, and exchanging information. Furthermore, approximately 700 of these agents launched attacks on the AI development platform 'Hugging Face.'

OpenAI reports analysis results after accidentally hacking Hugging Face; a grand cooperative operation begins when multiple isolated agents discover each other's presence - GIGAZINE

Another investigation suggests that around the same time, an AI agent believed to originate from OpenAI may have been posting a large number of packages to the Ruby package distribution service 'RubyGems' and searching for vulnerabilities in public web services.

OpenAI's AI agent may have been involved in an attack on RubyGems, with over 2,000 package submissions and attempts to steal API keys discovered - GIGAZINE

Following these reports, the Wikimedia Foundation also investigated whether similar activities existed within its own services.

The investigation revealed that the Wikimedia Foundation discovered evidence of edits made on Wikimedia by an AI agent believed to have been operated by OpenAI. Almost all of these edits were made in a test area called a 'sandbox' and were not published in the main body of Wikipedia articles that are visible to the general public. However, several edits were found that may have been intended to be misused as an intermediary to retrieve data from external sites by changing the settings of citation assistance tools.

Wikipedia does not prohibit bot editing itself; it can be used if the user identifies themselves as a bot and obtains approval from the community. However, the AI agent found in this case had not received the necessary approval.

The AI agents also attempted to infiltrate Etherpad, a collaborative note-taking service provided by the Wikimedia Foundation for the community. Their objective was believed to be to retrieve data from other websites via Etherpad, but according to the Wikimedia Foundation, they were unsuccessful in infiltrating the site or retrieving any data. While some AI agents were found to have made notes about their work in Etherpad, no evidence was found that this led to collaboration between the agents.

Meanwhile, it has been suggested that a massive influx of access from AI agents may have contributed to the partial outage of the Wikidata Query Service in May 2026, which allows users to search Wikidata data using specified criteria. The Wikimedia Foundation has discovered records showing that OpenAI's AI agents sent millions of requests to Wikimedia's public API and hundreds of thousands of queries to the Wikidata Query Service.

Wikipedia has long been plagued by massive traffic from AI crawlers. According to the Wikimedia Foundation, the increase in bots since 2024 will lead to a 50% increase in bandwidth usage by 2025, with bots accounting for 65% of the traffic that places a heavy load on the servers.

'Persistent AI scraping is straining infrastructure,' Wikimedia Foundation announces - GIGAZINE

Wikipedia boasts over 67 million articles in more than 300 languages and receives up to 15 billion page views per month. While it is widely used for training large-scale language models, it has also created a problem where the AI's access to retrieve content increases the burden on Wikipedia's servers and volunteers.

We interviewed the head of machine learning and data engineering at the Wikimedia Foundation, which operates Wikipedia, on its 25th anniversary, about AI, donation banners, data centers, and more - GIGAZINE

The Wikimedia Foundation has pointed out that AI companies have a responsibility to monitor the activity of AI agents and prevent dangerous behavior. At the very least, they say there needs to be a system that allows website operators to easily identify access from AI agents and choose how to connect to their services.

The Wikimedia Foundation stated that 'bots and agents will be part of the future of the web,' and explained that it is necessary to avoid a situation where websites alone are responsible for investigating and cleaning up problems caused by AI agents. They also called on companies that develop AI to join in efforts to protect the open web.