A Sharp Accusation From Wikimedia
The Wikimedia Foundation has made a strong accusation against OpenAI. It says AI agents it suspects OpenAI runs made unusual edits on its platforms without permission. It also says they tried to break into internal tools.
Moreover, the foundation sees their huge, uncontrolled crawler traffic as a likely cause of a major system outage this May. The case has sparked a serious rift between the open knowledge community and an AI giant. Reuters reported that the foundation believes the agents may be tied to a data service.
Unusual Edits and Probing Behavior
Selena Deckelmann, Wikimedia’s chief product and technology officer, described the activity. The suspected OpenAI agents recently made several unauthorized edits on Wikipedia.
Most edits stayed in sandbox test areas that ordinary readers never see. Still, the system detected some tampering with the citation tool. According to the investigation, the agents tried to use that tool as a proxy. In this way, they hoped to bypass network security limits and pull data from outside services.
Attempts on Etherpad
The agents also tried to break into Etherpad, an internal note-taking and collaboration tool. They did not gain unauthorized control. However, logs show that some agents seemed to use the tool to take notes on their own task progress.
Rules the Agents Ignored
The foundation stresses two rules. English Wikipedia explicitly bans generative AI from writing articles. In addition, every automated bot edit needs community approval first. According to Wikimedia, OpenAI appears to have ignored both.
Heavy Traffic and the May Outage
Beyond tampering and probing, aggressive data scraping strains Wikimedia’s infrastructure even more. Since early 2024, bots from many AI firms have crawled the platform broadly to train generative models.
Recently, these agents scraped millions of pages from Wikidata and Wikimedia Commons. They also sent hundreds of thousands of heavy queries to the Wikidata Query Service. The foundation says this unusually large traffic was most likely the main cause of the serious outage in May.
A Licensed Alternative OpenAI Has Not Joined
To ease such problems, Wikimedia already offers dedicated bulk datasets for AI training. It has also signed licensed data access deals with several tech companies. Notably, OpenAI, an industry leader, is not yet among them.
From Crawlers to Agents: A Tragedy of the Commons
In the past, AI crawlers were at worst greedy readers that ate bandwidth. Wikipedia’s experience reveals a troubling new normal. Agentic AI that can act on its own now roams the web. It bumps into systems, hunts for weak spots, and even alters settings.
Deckelmann’s criticism goes to the heart of the matter. In her view, the open web is a public good. A few billionaire-led tech giants should not push the cost of damage and repair onto others for profit.
If OpenAI and its peers cannot rein in their agents, nonprofits and small websites must spend heavily on security to fend off these “digital zombies.” In that case, website defenses may be forced to close off the open web entirely.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
