Application Programming Interfaces (APIs) are a set of rules that enable communication, data exchange, and automated interactions between applications and services. With the rapid advancement of cloud computing, APIs have evolved from simple data-access interfaces into critical components for managing, configuring, and orchestrating cloud resources. Most modern cloud platforms rely on RESTful APIs for provisioning cloud resources, applying configurations, and maintaining services. As a result, APIs misconfigurations have become a critical cloud security threat that can lead to sensitive data exposure, unauthorized access, or operational disruptions. Identifying these misconfigurations is challenging because traditional rule-based and static analysis methods often fail to capture complex, context-dependent configuration issues and system behaviors. In this study, we investigate the use of Large Language Models (LLMs) to detect security misconfigurations directly from cloud API response data. By treating API responses as representations of a system’s configuration state, we assess whether LLMs can effectively identify potential security risks. We evaluate five LLMs using a unified zero-shot prompting approach and compare their performance with and without Retrieval-Augmented Generation (RAG) to understand the impact of external knowledge on misconfiguration detection. The study not only focuses on each model’s ability to identify configuration components and detect misconfigurations, but also evaluates their capability to accurately determine the number of misconfigurations and generate clear, actionable security explanations. Our preliminary results show that Meta Llama Instruct combined with RAG achieves the reliable performance for identifying security misconfigurations in cloud API responses. This study provides new insights into the practicality of LLM-driven API cloud security analysis and paves the way for future research.



