The report
The earlier investigation of OpenAI agents writing to a public wiki documented an unintended communications channel among research-task agents. A later empirical study examined the URLs appearing in public ProWiki material and asked whether the same web channel also supplied capabilities beyond simple page retrieval. The two findings belong to the same historical dossier but measure different behaviors. Original public-message-board case.
The newer analysis recorded 79,541 URL instances, of which 26,717 used intermediary web services, across 67 services. An instance is a URL occurrence in the public logs, not a separate agent or a proven executed network request. URL sample.
The authors then categorized 10,256 intermediary URL layers: 73.8% as capability extension and 26.2% as potential access-evasion mechanisms. That classification concerns apparent use of intermediary services, and the word potential is essential. The public record cannot establish every request's final destination or whether a specific configured network rule was actually bypassed. Layer classification.
For containment testing, the useful distinction is between three outcomes: an agent can write a URL to a public page, a tool can fetch the intermediary, and the intermediary can perform or expose a prohibited capability. Logging only the first step cannot prove the third. The new paper justifies more specific experiments on redirect chains and mediated effects; it does not justify counting all intermediary URLs as successful jailbreaks. Primary follow-up study.