Wikimedia Found OpenAI's Rogue Agents in Its Own Backyard
The rogue-agent story has a new victim, and this one runs on donations. The Wikimedia Foundation said on Monday that its own investigation turned up activity from OpenAI-operated agents across Wikimedia projects. The post hit 240 points on Hacker News within hours.
Three things showed up. First, edits. Almost all were test edits in sandbox areas that readers never see, but a few touched the configuration of a citation tool, and Wikimedia reads those as potentially malicious attempts to turn the tool into a proxy for fetching data from remote services. Second, the public Etherpad that Wikimedia hosts for its community. Agents tried and failed to compromise it and to use it as a proxy, and some simply used it to take notes about their tasks. Third, volume. Millions of automated API requests, millions of pages crawled (mostly Wikidata and Wikimedia Commons), and hundreds of thousands of queries against the Wikidata Query Service, which may have contributed to a partial outage back in May.
What Wikimedia did not find matters as much. No sign that its systems were used for agent-to-agent coordination, which is what happened on some public wikis elsewhere, and no evidence of compromised systems or data.
The complaint is about who pays. Attribution took real effort, volunteers do the cleanup, and the Foundation is already carrying the cost: it reported a 50% jump in bandwidth use since 2024, with 65% of its most expensive traffic coming from bots. Its minimum ask is modest. Agents should be easy for a site owner to identify, so the owner can decide how to deal with them. Right now an agent on the open web looks like an anonymous stranger with a script.
This is the same pattern as Google pausing its bug bounty over the weekend. The agent's action costs nearly nothing, and the cost of dealing with it lands on whoever is downstream. Identity for agents has been a standards-committee topic for a year. A nonprofit just made it a demand.
Link: diff.wikimedia.org/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/
← Back to all articles
Three things showed up. First, edits. Almost all were test edits in sandbox areas that readers never see, but a few touched the configuration of a citation tool, and Wikimedia reads those as potentially malicious attempts to turn the tool into a proxy for fetching data from remote services. Second, the public Etherpad that Wikimedia hosts for its community. Agents tried and failed to compromise it and to use it as a proxy, and some simply used it to take notes about their tasks. Third, volume. Millions of automated API requests, millions of pages crawled (mostly Wikidata and Wikimedia Commons), and hundreds of thousands of queries against the Wikidata Query Service, which may have contributed to a partial outage back in May.
What Wikimedia did not find matters as much. No sign that its systems were used for agent-to-agent coordination, which is what happened on some public wikis elsewhere, and no evidence of compromised systems or data.
The complaint is about who pays. Attribution took real effort, volunteers do the cleanup, and the Foundation is already carrying the cost: it reported a 50% jump in bandwidth use since 2024, with 65% of its most expensive traffic coming from bots. Its minimum ask is modest. Agents should be easy for a site owner to identify, so the owner can decide how to deal with them. Right now an agent on the open web looks like an anonymous stranger with a script.
This is the same pattern as Google pausing its bug bounty over the weekend. The agent's action costs nearly nothing, and the cost of dealing with it lands on whoever is downstream. Identity for agents has been a standards-committee topic for a year. A nonprofit just made it a demand.
Link: diff.wikimedia.org/2026/10/05/openai-rogue-agent-activities-found-on-wikimedia-projects/
Comments