Identity and permissions aren’t enough to govern AI agent behavior

Box CISO Heather Ceylan argues that traditional identity and permission controls are insufficient for autonomous AI agents because they restrict what agents can access, not what they do afterward. As agents operate at machine speed and explore all available permissions, stale access and misconfigurations can quickly turn into unintended enterprise actions.

Enterprise AI security is shifting from access management toward runtime governance, where organizations monitor and constrain agent behavior during execution. This reflects a broader trend: as AI agents gain autonomy, companies will need layered controls that combine least-privilege access, tool-use limits, and continuous oversight.