LLM-Based Automated Mitigation and Assurance Case Generation Against Threats to AI Systems

Artificial intelligence (AI)-integrated systems face unique threats, vulnerabilities, and risks compared to conventional systems. Consequently, implementing security measures for AI systems in organizations is paramount. However, the security database, Adversarial Threat Landscape for Artificial-Intelligence Systems (ATLAS) has deficiencies in addressing and mitigating these concerns. Furthermore, while the application of large language models (LLMs) presents promising prospects in this field, thorough evaluations of their effectiveness have yet to be performed. In this study, we propose an LLM-based fully automated method for generating mitigations against AI systems threats and validating these mitigations using assurance cases. Furthermore, it presents innovative automated techniques for generating assurance cases, encompassing the dynamic collection of evidence, validation of effectiveness through these data, and solution nodes structured based on Toulmin's theory. We validated it using the method on 46 entries in ATLAS, producing mitigations and their corresponding assurance cases. To explore the potential superiority of the proposed method, we evaluated the mitigations based on information theory and assessed the assurance cases based on their completeness. The results indicate that the proposed method has the potential to generate more risk-specific mitigations, while the completeness of the produced assurance cases improved by up to 88.4% through each step of the proposed method.