New research shows enterprises granting excessive access to AI agents experience 4.5 times more security incidents. Traditional IAM does not work for autonomous systems that make dynamic access decisions. This article covers the identity gap, real incident patterns, and practical least-privilege patterns for AI agents.