ENSURING SECURE OPERATING OF VOICE AGENTS BASED ON THE LARGE LANGUAGE MODELS

Abstract

The rapid deployment of voice agents based on LLM (Large Language Models) in corporate systems creates new challenges for the software architecture and information system security. Traditional development approaches do not take into account the specifics of LLM integration as a component that can be compromised by attackers through prompt injections and other attacks at the natural language processing level. The article presents an architectural approach to secure voice agent integration based on the principle of separation of the basic business rules from the logic of interaction with language models. The proposed architecture creates a multi-level security system via 1. the implementation of the principle of least privilege for LLM components: 2. independent validation of business operations at the domain service level; 3. the usage of a large language model as a mediator between the user and critical systems. The research demonstrates the effectiveness of the architectural approach in countering typical attacks, including prompt injections to bypass authorisation, sensitive data leakage, unauthorised operations via function calling, and violation of business process integrity. The mediator architecture greatly simplifies the understandability and manageability of the system due to a clear division of roles between components, which increases the system predictability, and enables the use of traditional validation and audit methods. The architectural isolation of LLM from critical operations allows organisations to implement advanced AI technologies without a major overhaul of existing security systems, ensuring resilience to new vulnerabilities and flexibility in choosing of the technological solutions. The results of the research are of practical importance for the development of corporate systems with voice agent integration. Also, the resultscontribute to the forming of secure practices for the usage of large language models in applications of critical importance.