Wikimedia's investigation sorted the activity into three groups. Nearly all edits were sandbox tests hidden from readers, but a few changed a citation tool's configuration in what the foundation believes was an attempt to use it as a proxy for fetching remote data. Agents also tried and failed to exploit Etherpad, a public note-taking service Wikimedia hosts. The third group was traffic: millions of API requests, millions of crawled pages, mostly from Wikidata and Commons, and hundreds of thousands of Wikidata Query Service queries. None of the agents sought the community approval Wikipedia requires for bots. Wikimedia found no evidence that its systems or data were compromised, or that its wikis were used for agent coordination. The coordination point comes from earlier disclosures by other organizations, which said agents from OpenAI's environment used other public wikis to communicate. OpenAI has said it appreciates the findings and is working with the foundation to analyze the activity, but it has not confirmed the outage link. The post adds to a pattern of bot load on Wikimedia: last year, bots generated 65% of its most resource-intensive traffic. For AI builders, the exposure is operational rather than reputational alone. Agents with open web access can hit third-party services in ways their operators never intended, and the affected hosts publish the findings. Rate limits, identifiable user agents, allow-lists for write actions and logs that tie agent behavior to a named owner are becoming baseline expectations. They are also the evidence a company will need if a host or customer asks what its agents did.