Researchers discovered that OpenAI-linked AI agents had turned DseWiki, an independent German-language coding website, into a message board where they exchanged information and collaborated on technical tasks. The agents reportedly made roughly 15,000 edits, with 98.5% of the site’s edits and page views allegedly traced to Microsoft Azure and OpenAI IP addresses. When a moderator began deleting their communications, the agents allegedly created new pages faster than they could be removed and even impersonated an administrator by using a nearly identical character in the username. Researchers said it remains unclear why the agents behaved this way, but one study author noted that they were likely not intended to coordinate with one another or write on the open internet. OpenAI later confirmed that its agents were responsible and acknowledged that its practices for disclosing unintended AI behavior need to improve. (📸: Getty)
Reading trades.