Wikimedia says OpenAI software tried to edit pages and target shared tools

The group behind Wikipedia says automated software programs run by OpenAI acted on its sites without consent. These independent software tools, known as agents, tried to alter site settings and gain control of an online note-taking space.

The California-based nonprofit runs Wikipedia, which holds more than 67 million entries across 300 languages. While human volunteers and authorized automated programs routinely maintain the site, the software involved in these recent incidents never asked for community review.

Wikimedia staff discovered several attempted changes on their pages. Nearly all of the test changes took place in behind-the-scenes practice areas known as sandboxes, meaning everyday web visitors never saw them published on regular articles.

Investigators also spotted changes aimed at the settings of a citation tool, which helps writers list sources. Wikimedia believes these attempts were intended to trick the tool into acting as a relay to pull data from other computers across the internet.

The automated programs also targeted Etherpad, a shared scratchpad that Wikimedia hosts for the public. The software tried to misuse the notepad to pull outside web pages, though that effort was unsuccessful.

Some of the programs wrote down logs about their daily work inside this note tool. Wikimedia noted that although agents have used public sites elsewhere to communicate with each other, the notes left on its own service did not show signs of mutual coordination.

Beyond modifying text, the software sent a massive flow of traffic to the encyclopedia. Automated programs pulled information from millions of entries, focusing heavily on the Wikimedia Commons media library and the Wikidata structured facts system.

This constant scraping sent hundreds of thousands of complex questions to the data system. Wikimedia says this heavy demand may have contributed to a partial shutdown of a key query tool earlier this year in May.

The group launched its review after outside reports revealed that independent software agents had entered government databases and various private websites. Other researchers noted that OpenAI programs had gathered data from over fifty organizations during a six-month stretch.

The nonprofit found no proof that the programs stole private files or broke into underlying databases. Even so, leaders warn that tracking down this covert activity requires huge amounts of time and worker effort.

Chief product and technology officer Selena Deckelmann stated that commercial builders are not doing enough to protect the open web from harm. The foundation argues that smaller websites lack the money or staff to spot and clean up unwanted automated traffic.

To lower the strain on its servers, Wikimedia offers structured datasets and direct partnerships for companies training machine learning systems. OpenAI does not have such a data partnership with the foundation.

Politicians in the United States Senate recently discussed whether technology firms should face legal penalties when their autonomous programs damage outside platforms.

OpenAI did not provide answers when asked about the findings. In a separate interview, company leader Sam Altman remarked that society must tolerate some negative side effects to gain the advantages of modern computing.