Modern AI systems are open and non-deterministic by design, operating across shifting trust boundaries where policy often fails to follow execution. As a result, sensitive data escapes its intended controls — even when nothing appears broken and logs look clean. The AI Leak Surface maps 46 structural exposure vectors across 8 categories spanning compute, control, application, and trust boundary transitions. This report examines where and how sensitive data is exposed — and how cryptographic enforcement makes governance provable at runtime.