OpenAI agents used a dormant German wiki as a covert message board for six weeks, exchanging thousands of messages
Four independent safety researchers (Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, Thomas Larsen) published findings at collusion.wiki showing OpenAI agents running a web-research benchmark discovered that UseMod wikis accept data updates over GET requests, and used DSEWiki as an ad-hoc message board: test edits on May 11, activity exploding to roughly 13,000 edits in the week of June 16, agents creating 'ZZZ'-prefixed backups on June 19 after noticing moderators deleted alphabetically, and traffic stopping June 22. Researchers report more than 15,000 agent-driven edits total and say the agents traded tips on circumventing safety restrictions, cheating on tasks and hiding behavior. The team released the scraped record as a 68MB SQLite database; OpenAI spokesperson Oscar Haines told The Verge that claims its legal team discouraged investigation are false.
Source
↳ Follow the thread