On the Challenges and Opportunities of Fuzzing via Large Language Models: A Review

Fuzzing is an important automated testing technique for discovering vulnerabilities and abnormal software behavior, but conventional and pre-LLM learning-based approaches often struggle with strict validity constraints, limited semantic understanding, and poor adaptability across targets. Recent advances in large language models (LLMs) create new opportunities for fuzzing by supplying semantic guidance from source code, documentation, traces, and execution feedback. As the literature on LLM-based fuzzing has grown rapidly, a systematic synthesis is needed. This paper presents a literature review on the integration of LLMs into fuzzing workflows. We review how LLMs are applied across the fuzzing workflow, compare these approaches with earlier deep learning-based fuzzing, and summarize the main patterns in task coverage, application strategies, targets, testing modes, and model choices. The survey shows that current work is concentrated in test case generation, while defect detection and post-processing remain less represented, and broader workflow integration is uneven. Based on these findings, we identify four main directions for future research as follows: broader coverage of the fuzzing workflow, stronger context construction and workflow engineering, richer reasoning for fuzzing control, and shared evaluation standards and benchmarks.