Secure deployment practices for large language model applications

Abstract

This thesis identifies security threats associated with deploying LLM-based applications and provides organizations with effective strategies for mitigating these threats while maintaining the functional integrity and utility of their AI systems. In order to achieve this, prior literature has been systematically reviewed, current pressing vulnerabilities affecting LLM applications have been analyzed, and implemented mitigation strategies have been empirically evaluated.

After systematically analyzing previous literature, a handful of mitigation strategies for significant vulnerabilities, applicable for the majority of LLM applications, were selected for further examination. The evaluation results imply that the implemented security measures are highly effective in preventing or mitigating the related vulnerabilities.