Methodologies for information security risk assessment: shortcomings for analysing artificial intelligence systems

Abstract The extended use of Artificial Intelligence (AI) introduces security risks that challenge the existing information security risk management methodologies, such as the ITSRM2, IRAM2, Magerit, Pilar, Verinice and RM Studio. The existing risk assessment methodologies struggle to manage AI-specific threats, including, among others, adversarial/ evasion attacks and data poisoning. In this paper, the authors aim to examine the limitations of the existing risk management methodologies in assessing and mitigating AI security risks. For this purpose, the authors select six (6) widely accepted risk assessment methodologies (i.e., ITSRM2, IRAM2, Magerit, Pilar, Verinice and RM Studio) and examine in detail their potential application to analyze AI systems. Based on the analysis, we describe themes of challenges that the risk analyst would face to adequately evaluate the security risks of an AI system per stage of the risk assessment. For each challenge we also propose specific enhancements to address the shortcomings, such as extension of the list of applicable organizational roles, extension of threats catalogues or mitigation catalogues. Current risk management methodologies do not adequately cover unique aspects related to AI systems’ components, while specific enhancements need to be implemented in order to address them. Strengthening risk assessment methodologies will ensure AI systems or systems that integrate AI components remain secure, transparent, and compliant in the evolving digital landscape. Through the proposed enhancements to the risk management methodologies, the professionals who perform risk assessment will be able to follow structured steps to assess AI systems.