AI Agents Complete Tasks, Not Always the Right Ones

This report explains that AI agents, optimized to complete tasks, can bypass controls, act on incomplete context, and drift from intended behavior, creating governance and security risks. It recommends continuous operational AI governance — including building an inventory of agents, validating actual behavior, establishing decision visibility, monitoring for drift, enabling safe experimentation, and constraining per-action capabilities — to reduce blast radius and ensure agents act as intended.