Glossary
Runtime governance
Controlling what AI agents do while they run, one action at a time, instead of only checking what a model says or reviewing an agent before launch.
OWASP's Top 10 for Agentic Applications describes the controls: at run time, validate both the user's intent and the agent's before goal-changing or high-impact actions; put a policy enforcement point in front of every tool call; give each agent its own short-lived, narrowly scoped credentials; cap cost and rate; keep immutable logs of every tool call; and be able to disable a rogue agent instantly with a kill switch and credential revocation.