On Cyber Risks of Generative Artificial Intelligence

Abstract

This article is devoted to an overview of the risks of generative models of Artificial Intelligence. The rapid development of large language models has seriously increased attention to the security of Artificial Intelligence models. From a practical point of view, in this case, we are talking about the security of deep learning models. Large language models are susceptible to poisoning attacks, evasion attacks, attacks aimed at extracting training data, etc. But, at the same time, new attacks appear that are related specifically to the generated content. Moreover, the latter constitute an obvious majority. Therefore, recently many works have appeared that try to systematize all the risks of generative models. For example, OWASP and NIST are engaged in this. A complete taxonomy of generative AI risks should serve as a basis for building testing systems for generative models. This paper provides an overview of generative AI risk specifications outlined by OWASP, the NIST profile, and the MIT risk repository. The purpose of such specifications is to create a base for testing generative models and tools intended for AI Red Teams.