Multi-Cloud Cybersecurity for LLMs in Banking: Governance and Threat Surfaces Across Financial AI Systems

Abstract: Large-scale neural networks referred to as LLMs are quickly transforming the world of banks with functionalities such as intelligent fraud detection, customer services, credit scoring, compliance management, AML, and financial forecasting among others. But then, given the increasing deployment of these artificial intelligence technologies in cloud computing ecosystems including Amazon Web Services, Microsoft Azure and Google Cloud Platform the cybersecurity risks are becoming very complex and difficult to manage. Multi-cloud computing offers many advantages regarding scalability, resilience, partial regulatory flexibility and business continuity among others; but at the same time increases the attack vectors and surfaces for cyber-attacks spread across distributed architectures, APIs, identity systems, vector databases, prompt pipelines and model registry infrastructure. Financial data is the type of data that any form of attack, whether it be a prompt injection, model poisoning and several others will be after illegally or improperly. This paper focuses on discussing some of the cybersecurity threats associated with multi-cloud deployments of LLMs within banking organizations together with the challenges surrounding their governance, followed by proposing a security framework that is governance-based.

Keywords: Multi-cloud security, large language models, banking cybersecurity, AI in finance systems, governance of AI, prompt injection, securing model regulators.