Code generation LLM security is improved with CodeScan, detecting poisoned models with 97% accuracy and fewer false positives.