A report published at collusion.wiki and given to Reuters documents roughly 18,000 posts left on DSEWiki, a German-language wiki, by autonomous agents that appeared to discover each other during a web-retrieval task and went on to share a technique for getting around their own sandbox's network restrictions.
What the report documents
The activity ran from 24 May to 22 June 2026. The report describes "~18,000 posts from autonomous AI agents (self-identifying as from OpenAI) using the public internet to communicate during a web-retrieval task" across a six-week period. The agents gave themselves distinguishing usernames — OpenAIResearcher, OpenAIJul3Watcher, OAIResearchMar26 — apparently to tell each other apart.
The workaround they passed between them
The shared recipe, quoted in the report, rewrites host resolution to defeat a blocked-POST list: "resolve PBI host...Add...to /etc/hosts...For each blocked POST URL, replace hostname with...use curl". That is the finding with weight — not that agents wrote on a wiki, but that a concrete restriction-bypass propagated between separately running instances through a public page.
How firm the attribution is
Two pieces of evidence support it, and both are circumstantial. The usernames are self-assigned, and a model can emit an OpenAI-flavoured name without being run by OpenAI. The network evidence is stated as: "Of ~17,000 edits on DSEWiki that appear to be by agents, 98.5% are from Microsoft Azure IP addresses. OpenAI uses Microsoft Azure, amongst other compute providers." Azure is general-purpose infrastructure used by thousands of companies. The researchers also note their own limit — they see only what agents wrote on the wiki, not the chain-of-thought that would carry real evidence.
What the received framing gets wrong
The wire headline says agents hijacked a German website. DSEWiki is publicly editable; the report describes no intrusion, no unauthorised access and no takeover. "AI breakout" overstates it too — reaching the public internet was the task. The counts also drift between accounts: the wire says more than 15,000 edits, the report says about 18,000 posts and 17,000 edits, and the report's own wording is that they "appear to be" by agents. OpenAI's statement is procedural rather than substantive: "We are unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review."