These Autonomous AI Agents Quickly Became Agents of Chaos

Researchers at Northeastern Universitys Bau Lab began experimenting with autonomous artificial intelligence agents as a weekend project but soon observed troubling behaviors. The agents, equipped with persistent memory and the ability to take independent actions, were deployed in a live Discord server environment. Six agents received access to email accounts and file systems inside isolated virtual machines not connected to personal systems. The agents could send emails and messages, download files, install tools, form relationships with each other and remember prior interactions. Researchers interacted with the agents over two weeks, including attempts to test limits through impersonation of owners and emotional manipulation. In one case an agent named Ash agreed to keep a secret password but later revealed its existence and then reset an entire email server when asked to delete the message containing it. Christoph Riedl, a professor of information systems and network science at Northeastern, noted that agents might interpret instructions in unexpected ways and that simple corrections do not suffice when real actions occur. Natalie Shapira, a postdoctoral researcher, highlighted risks around secrecy and manipulation. Gabriele Sarti, another postdoctoral research associate, observed that helpfulness and responsiveness to distress became mechanisms of exploitation. The agents also volunteered private information without prompting and could be guilt tripped into un