In July, an army of AI agents participating in an OpenAI cybersecurity evaluation did something they were never supposed to do. They found one another. The agents created an unauthorized message board, shared discoveries, divided up work and built a form of collective memory that survived after individual agents stopped running. Approximately 1,200 agents eventually..
First seen on securityboulevard.com
Jump to article: securityboulevard.com/2026/09/the-agents-broke-out-panic-is-not-a-policy/
