Securing the LLM Backend: Mitigating Emerging Threats and Ensuring Data Privacy for AI-Powered Financial Applications on AWS

Abstract

Large Language Models (LLMs) have a significant impact on the financial technology (fintech) industry as the rapid growth of these automation types has simplified the processes of automation, customer interaction, and predictive analytics. Still, such breakthroughs bring a series of cybersecurity and privacy risk factors that compromise the data integrity, regulatory adherence, and institutional trust. This paper analyzes the security risk factors that are inherent to LLM backbend’s that are deployed in AI-based financial applications on Amazon Web Services (AWS). Using a hybrid-methodology combining qualitative threat modeling with quantitative analysis of AWS-specific security settings, we use the STRIDE and MITRE ATT&CK framework to name key vulnerabilities, such as prompt injection, data exfiltration, model inversion, and privilege escalation. The next steps that we take are to assess the effectiveness of the AWS-native mitigation aspects like in-use encryption, granular Identity and Access Control (IAM) controls, network segregation, and continuous auditing. The findings show that the integration of the governance of LLM with the cloud security architecture of AWS significantly increases data confidentiality and contributes to the international financial regulation, in particular, GDPR and PCI-DSS. This paper introduces a security-by-design model of LLM backends, where explainability and data minimization as well as proactive monitoring are crucial. Therefore, the paper highlights the critical importance of safe AI-cloud integration in privacy, robustness, and trust protection in financial ecosystems