Tool-Guided Retrieval-Augmented Repair for Securing LLM-Generated C Code

This paper presents an analysis-and-repair workflow for LLM-generated C code that combines compiler diagnostics, CodeQL static analysis, KLEE symbolic execution, and retrieval of prior repair patterns. Across 5,000 C programming tasks involving embedded-relevant vulnerabilities, the workflow reduced CodeLlama 7B’s security defect rate from 49% to 19% and lowered total CodeQL errors from 15,088 to 2,463, an 83.7% reduction. For DeepSeek Coder 1.3B, compilation failures fell from 42% to 22%, while security defects dropped from 35% to 15%.