We propose a distillation framework that transfers secure coding expertise into a compact model trained using expert demonstrations, static-analysis labels, and secure patterns. The distilled model (7B parameters) achieves 83% security compliance, nearly matching GPT-4’s 89% performance while requiring significantly less compute. Evaluated across CWE-20, CWE-78, CWE-89, the distilled model reduces injection vulnerabilities by 62% relative to baseline models. This work demonstrates that targeted distillation can teach smaller models to respect security best practices.