Systematic Testing of Security-Related Vulnerabilities in LLM-Based Applications

Large Language Models (LLMs) have emerged as transformative tools in natural language understanding and generation. They possess billions of parameters, which enable them to generate coherent and contextually rich text [1]. These capabilities have made LLMs vital in domains such as customer service, content creation, and programming assistance [2], [3]. However, these advances come with significant risks. For example, a recent study showed that LLM responses contain private or sensitive information accidentally exposed during training [4], [5]. Furthermore, adversarial attacks have been shown to reduce system accuracy by as much as under controlled conditions [6]. A high-profile example is the misuse of LLMs to generate biased or harmful text when manipulated through adversarial prompts [7].