LLMSQLi: A Black-Box Web SQLi Detection Tool Based on Large Language Model

Black box detection tools of SQL injection vulnerabilities simulate real-world web attack scenarios, making it essential for evaluating SQLi risks in actual web applications. However, current black-box approaches depend on predefined rules for SQLi vulnerability detection, which limits both their efficiency and accuracy. In this paper, we propose a black-box SQLi detection tool based on large language multi-agent, LLMSQLi, which uses the context understanding and reasoning capabilities of large language models and the cooperative division of labor mode of multi-agent to generate payloads customized for test targets and efficiently detect SQL injection vulnerabilities in Web programs. Drawing inspiration from real-world teams of security experts, LLMSQLi simulates the step-by-step process of human experts in testing tasks through the LLM Muti-Agent collaboration model. We ran experiments on SQLiMicroBenchmark to compare the performance of LLMSQLi and two of the most advanced black-box SQLi testing tools. Experiments show that LLMSQLi successfully detected all 15 targets and outperformed other tools.