Unmasking Threat Vectors in Large Language Models

AI-driven defect detection is increasingly essential as software systems grow in complexity, multilinguality, and scale. This chapter presents a comparative evaluation of traditional machine learning models and large language models for detecting structural and semantic defects across multiple programming languages. A multilingual dataset is used to assess Random Forest, SVM, XGBoost, and a prompt-based CodeT5 simulation, revealing the limitations of feature-engineered approaches and the superior semantic reasoning of LLMs. Practical integration scenarios in CI/CD, IDEs, and DevSecOps workflows are examined, followed by key challenges and emerging research directions for scalable, explainable, and automated defect detection.