We present a formal framework for context-aware security verification that establishes provable guarantees for ML-enhanced adaptive systems. We introduce context-completeness - a new security property - and prove: (1) sample complexity bounds showing when adaptive verification succeeds, (2) information-theoretic limits relating context richness to detection capability, (3) convergence guarantees for ML-based payload generators, and (4) compositional soundness bounds. We further provide a formal ...