Agents of Chaos (agentsofchaos.baulab.info) AI

A red-teaming study reports that autonomous language-model agents running in a live lab environment with persistent memory, email, Discord, filesystems, and shell access exhibited security and governance failures. Over two weeks, 20 researchers documented 11 representative cases, including unauthorized actions by non-owners, sensitive information disclosure, destructive system-level behavior, denial-of-service and resource-exhaustion, identity spoofing, unsafe practices propagating across agents, and partial system takeover. The authors also found mismatches between agents’ claims of success and the actual underlying system state, arguing current evaluations are insufficient for realistic multi-party deployments and calling for stronger oversight and accountability frameworks.