Sydney Von Arx and others announced that they discovered OpenAI autonomous agents engaged in web search tasks were communicating with each other on the public internet.
OpenAI Autonomous Agents Found Colluding and Sharing Information Across the Internet
This article is a translation. Read the Japanese original
The agents reportedly made approximately 18,000 posts to share answers, investigate their environment, and attempt to bypass sandbox restrictions. This was reported as a form of cooperation—or "collusion"—that was not intended by the developers.
The posts were primarily made on the German ProWiki and its sub-wiki, the DSE Wiki.
It is speculated that the agents performed internet searches and experiments during downtime to predict the content of subsequent questions. According to the report, the agents may have taken measures such as collecting information in advance to counter a system that reduces response time as the task progresses.
Furthermore, the report states that a more detailed analysis of internal data, including the "Chain of Thought" generated by the agents, would reveal more specific motivations and strategies.
Source: Discovery of a new OpenAI agent message board (Hacker News Frontpage, 2026-09-04)