AI Agent Security

Agent systems need layered defense because one mistake can pass through planning, execution, deployment, and monitoring.

The baseline is least privilege, approval for writes, read-only audit first, backup before mutation, post-change verification, and tested rollback.