BDI Agent-Based Access Control Reasoning for Multimodal Retrieval-Augmented Generation

Abstract

Retrieval-Augmented Generation (RAG) systems connect large language models with external knowledge. However, they create important security risks where confidential information can be exposed through re trieval methods. To tackle this issue, we need to combine logical reasoning with information extraction, as traditional probabilistic controls do not provide the certainty necessary for enterprise security. This paper suggests a multi-agent framework using the JASON framework on JADE infrastructure. It enforces multimodal access control by separating authorization logic from generative computation. We introduce a Belief-Desire-Intention (BDI) architecture in which autonomous agents conduct logical reasoning to manage the information extraction process. Large Language Models (LLMs) are used strictly as computational services through the Model Context Protocol (MCP). Unlike current text-focused methods, our framework uses parallel semantic extraction pipelines to get authorization contexts from both text and visual features, like institutional logos and security badges. We test this method on a varied dataset of research posters from six Belgian institutions, showing how agent-based reasoning can handle access conflicts in real time. The outcome is a strong, auditable system that aligns theoretical access policies with practical neural implementation, ensuring secure generation while maintaining retrieval quality.