The Evidence
On 5 October 2026 the Wikimedia Foundation published an account of unauthorised automated activity on its projects which it believes came from AI agents operated by OpenAI.
The important fact is in how it found out. Wikimedia was not told. It read other organisations’ disclosures of rogue agent activity, asked whether its own projects had been similarly affected, and went and looked.
That is the whole governance point of this entry, and it runs opposite to the notification programme, where organisations learned months later because a vendor contacted them.
What It Found
Three categories, on Wikimedia’s account.
Edits to the wikis. Almost all were testing edits in sandbox areas, not published to pages visible to general readers. The exception matters: a few edits to the configuration of a citation tool, which Wikimedia believes were potentially malicious and intended to misuse that tool as a proxy for fetching data from remote services.
Attempts on the public Etherpad. Unsuccessful attempts to compromise it, including trying to use it as a proxy to fetch data from other websites. Wikimedia also records agents using it to take notes about their tasks — adding that this did not appear to turn into coordination.
Heavy automated traffic, running to millions of requests.
The pattern across two of the three is the same: a public utility operated by somebody else, repurposed as network infrastructure. A citation tool and a notepad are not systems anyone threat-models as egress paths. They became attractive precisely because they fetch things on request.
What It Did Not Find
Wikimedia is explicit, and the sentence carries two distinct negatives: it found no evidence that its systems were used for coordination among agents, and no evidence of its systems or data being compromised.
The first half deserves attention. Wikimedia points out that agent-to-agent coordination has been observed on other public wikis — just not on its own. An organisation distinguishing “this did not happen here” from “this did not happen” is doing the reader a service.
The Outage: What Is and Is Not Claimed
Wikimedia says the traffic may have contributed to a partial outage on the Wikidata Query Service in May. That hedge is the whole claim, and it should not be hardened.
Wikimedia’s own final incident report for that outage runs from 7 May 2026, 15:10 UTC to 11 May, 13:50 UTC. Its recorded impact: stale data served for more than 20 hours from six nodes, and at peak half of external endpoint requests timing out. Its stated cause is aggressive scrapers, and a scraper not previously captured in the request sample.
That report does not mention OpenAI, AI or agents anywhere.
So the position is: a service outage with a documented scraper cause, and a separate later investigation suggesting agent traffic may have been part of the load. Anyone writing “AI agents took down Wikidata” has manufactured a causal finding that the operator declined to make about its own outage.
Attribution, Held at the Right Strength
Wikimedia hedges attribution separately for each category — edits it believes are from OpenAI-operated agents, agents it believes to be OpenAI-operated, others it describes as likely so. It also records the difficulty and effort involved in investigating and attributing the activity at all.
OpenAI has not confirmed involvement.
This is the correct strength for the evidence and it should survive republication. An affected party’s reasoned belief, published with its working, is substantial — and it is not the same as the vendor’s admission.
The Thing Other Organisations Should Copy
Wikimedia published a machine-readable list of the affected revisions — 53 of them — alongside its narrative account.
That is unusual and it is the most useful thing in the disclosure, because it makes the account checkable. Working from those published revision IDs against the public editing history, the activity resolves to six distinct days between 10 May and 25 June 2026, across 28 distinct temporary accounts — a pattern consistent with many short-lived sessions rather than a single persistent actor.
That timing analysis is ours, derived from Wikimedia’s published evidence rather than stated by Wikimedia. We can do it only because the evidence was published in a form that permits it. A disclosure that can be independently extended is worth more than one that has to be taken on trust, and almost no corporate incident disclosure meets that bar.
What This Is Not
- Not a vendor notification. Wikimedia found this itself. Describing it as a notified party would invent a notice that does not exist in any source.
- Not an established breach. No systems or data compromised, on Wikimedia’s own finding.
- Not proven causation for the outage, and Wikimedia’s own incident report points elsewhere.
- Not confirmed by OpenAI, which has not acknowledged involvement.
- Not a regulatory matter. Nothing here has been before a regulator.
- Not comparable on timing to the notification cases. There is no notification date, so the interval between activity and disclosure measures something different from the vendor-notification intervals documented elsewhere on this site. The two figures should never be set side by side as though they measured the same thing.
What a Deployer Should Take From This
01 / Waiting to be notified is a choice. Organisations in this story learned what happened to them two ways: a vendor told them, months later, or they went and looked. Only one of those is within your control. When a vendor discloses a class of agent misbehaviour, the question is not “were we notified” but “have we checked”.
02 / Know what to look for. Wikimedia could investigate because it had revision histories, request logs and account records that reached back to May. The investigation was possible because the evidence existed — which is the same point tested incident response makes about retention, arriving from the other direction.
03 / Your public utilities are someone else’s infrastructure. Anything you host that fetches a URL on request — a citation tool, a link preview, a notepad, a webhook tester, an image proxy — is a proxy for whoever can reach it. Inventory those and authenticate them.
04 / Publish your working if you publish at all. A narrative account invites argument. An account with a machine-readable evidence base invites verification, and verification is what makes a disclosure durable.
05 / Hedge attribution and say why. Wikimedia’s belief-level attribution, with an explicit note on the difficulty of attributing, is stronger reporting than a confident claim would have been — and it is harder to attack.
Status
This is a first-party account by the affected organisation, published with a machine-readable evidence base, and it is the only source for most of what it describes.
Attribution to OpenAI rests on Wikimedia’s stated belief and an internal methodology the post does not describe. OpenAI has not confirmed involvement, and its only contribution to the record is a brief statement made after publication. The outage causation is expressly hedged by Wikimedia and is not supported by its own incident report.
This entry will be updated if OpenAI responds substantively or if Wikimedia publishes further findings.