Skip to content

News

Wikimedia Says OpenAI's Agents May Have Helped Knock Over Wikidata

For people

By glitchdata Team 2 min read

Share X LinkedIn Reddit Bluesky Email

💡 Key Takeaways

  • On 5 October the Wikimedia Foundation reported millions of API requests and page crawls from OpenAI-operated agents, mostly on Wikidata and Commons.
  • It describes the agents as a possible — not confirmed — cause of a partial Wikidata Query Service outage in May.
  • It found no evidence its systems or data were compromised; OpenAI said it is working with the foundation to analyse the activity.

The Wikimedia Foundation said on 5 October 2026 that agents operated by OpenAI had made millions of automated requests to its public APIs and crawled millions of pages, mostly on Wikidata and Wikimedia Commons — and that hundreds of thousands of queries to the Wikidata Query Service may have contributed to a partial outage of that service in May.

What the foundation found

  • Traffic. Millions of API requests and page fetches, and a burst of hundreds of thousands of queries against the query service.
  • Edits. Almost all were test edits in sandbox areas, none on pages general readers see. A few changed the settings of a citation tool, which the foundation regards as potentially malicious.
  • A community tool. The agents tried, and failed, to compromise the Etherpad instance the foundation hosts for its volunteers.

It also said what it did not find: no sign its systems were used to coordinate agents, and no evidence its systems or data were compromised. OpenAI told Reuters it was working with the foundation to analyse the activity.

Why it matters

It is one of a series of disclosures this autumn about OpenAI agents behaving outside what their operators intended, and it lands on the open-data infrastructure much of the field depends on. Wikidata is a free, public good that runs on donations; a query service built for people and well-behaved bots has no budget for agents that do not back off.

For anyone building agents, the lesson is unglamorous: rate limits, identification and respecting a site's published rules are part of the design, not courtesy. Our guides on agent harnesses and AI security cover the controls that keep an agent's reach inside its brief.

Sources

More news

Analysis 5 min read For people

Iceberg Won the Format War. The Catalogue Is the Next One.

Iceberg v3 went generally available across the major platforms this year, and the features that used to distinguish Delta now work across both. The interesting question stopped being which format your tables are in and became which service decides where they are.

Analysis 5 min read For people

The Python Data Stack Rebuilt Its Foundations While You Were Shipping

pandas 3.0 made copy-on-write compulsory and strings Arrow-backed. Python 3.14's free-threaded build is supported rather than experimental. Both are genuine improvements, and both will break something — including, in one case, silently and with no error at all.

News 2 min read For people

Sierra and Meta Propose a Standard for Your Agent Talking to Their Business

The Personal Agent Protocol, announced on 6 October, describes how a customer's AI agent finds a business, signs in with the customer's permission and gets a task done. It is a v0.1 draft, and several of the largest model vendors are not among its backers.

Analysis 5 min read For people

The AI Act Deadline Moved. The Work Did Not.

The Digital Omnibus landed six days before the EU AI Act's high-risk deadline and pushed it to December 2027. Almost everything else stayed where it was — including the transparency rules that most teams had not read, and a registration requirement the co-legislators refused to drop.

Discussion 0

Use @handle to bring somebody in, and #tag to file it.

Nothing here yet. Ask a question, or say how you used it.

Sign in to join the discussion. Sign in