AutoBoo: Automating Fuzzing with LLM-Generated Templates

作者：Huan Le, Liming Zhu, Yan-bo Chen, Peiyu Liu, Liyang Hou, Jiaao Sheng · 年份：2025 · DOI：10.1109/eiecs67708.2025.11283482 · 研究领域：Software Testing and Debugging Techniques、Software Engineering Research、Web Application Security Vulnerabilities

Fuzzing is a widely used technique for discovering vulnerabilities in software systems, but it often relies on manually defined input templates, which can be time-consuming and error-prone. In this paper, we introduce AutoBoo, a novel approach that leverages Large Language Models (LLMs) to automatically generate input templates for Boofuzz, an open-source fuzzing framework. By integrating LLMs into the fuzzing process, AutoBoo automates the creation of diverse and syntactically correct input templates, eliminating the need for manual template definition. Our method uses LLMs to analyze the target system’s specifications, protocols, and previous input data to generate robust and dynamic templates that conform to the syntax of the system under test. These generated templates are then used in Boofuzz to perform effective fuzzing with minimal human intervention. We demonstrate the effectiveness of AutoBoo through experiments on several real-world applications, showing that it significantly increases fuzzing coverage, reduces setup time, and uncovers potential vulnerabilities that might otherwise be missed. This approach offers a significant advancement in automating the fuzzing process, making it more accessible and efficient.