Taxonomy of challenges influencing secure code generation using LLMs: An integrated approach of systematic literature review and F-AHP

The growing adoption of Large Language Models (LLMs) in software development has introduced new opportunities for automation but also significant concerns regarding code security and reliability. Ensuring the generation of secure code through AI-assisted tools has therefore become an important research challenge. The aim of this study is to identify, validate, and prioritize the key challenges that compromise security in code generated using LLMs. Moreover, a taxonomy of the identified challenges and their categories has been developed based on their relative importance, evaluated through the Fuzzy Analytic Hierarchy Process (F-AHP). In the first stage, a Systematic Literature Review (SLR) was conducted to extract potential challenges from 22 primary studies. In the second stage, the identified challenges were empirically validated through a questionnaire survey involving software security and AI practitioners. In the third stage, the F-AHP method was applied to evaluate and prioritize the significance of these challenges. The study reports thirteen major challenges, including training dataset quality, poor prompt structure, data poisoning, vulnerable dependencies, and adversarial prompt attacks. The proposed taxonomy provides a structured framework that assists both researchers and industry professionals in understanding, assessing, and mitigating the security risks associated with LLM-based code generation.