Summary

The OWASP Top 10 for LLM Applications has become the reference point for AI risk — ten ranked, testable failure modes, from prompt injection and sensitive information disclosure to excessive agency and unbounded consumption. It's the checklist boards ask about, auditors test against, and regulators increasingly cite. But naming a risk doesn't stop it. Below, each of the ten meets the AI Security Control Plane: what enforces it, how it's answered, and the outcome — allow, deny, mask or rewrite — decided inline.