Natural Language to SQL (NL2SQL) interfaces has made data far more accessible, but they have also opened up a new kind of weakness we call “Semantic Drift.” Because these systems are probabilistic, a Large Language Model (LLM) can produce SQL that is syntactically correctyet still breaks implicit privacy rules through hallucination or adversarial prompt injection. Conventional, deterministic security models are not built for this: they check identity, not semantic intent. To close that gap, we apply discrete combinatorial mathematics—specifically, dynamically generated Magic Squares—to build a positional symmetric cipher that functions as an adaptive Moving Target Defense (MTD). A Tri-State Decision Engine sits inside transparent JDBC middleware, balancing data integrity against active defense. When a query looks anomalous, the system responds with high-entropy obfuscation generated through the Siamese method, scrambling sensitive data on the fly and defeating frequency analysis in the process. In our evaluation, this approach caught 97.4% of adversarial injections while adding only 6.8ms of latency overhead, pointing toward a new cryptographic standard for zero-trust database connectivity in AI-driven systems.