Secure AI Integration: Evaluation of the Model Context Protocol(MCP) for Controlled Access for LLM-Based Chatbots : Design and evaluation of a security gateway for enterprise AI systems

Sammanfattning

This thesis investigates the security vulnerabilities that arise when large language models are integrated with internal enterprise systems via the Model Context Protocol(MCP), a protocol that acts as a bridge between AI models and external resources. Through an empirical analysis of two production grade enterprise systems, using staticcode analysis, dynamic testing, and developer interviews, several critical security vulnerabilities were identified, including token pass-through, exposure of authenticationcredentials, missing endpoint authentication, and insufficient input validation. Theresults confirm that current MCP implementations systematically prioritize rapid integration over security robustness, leading to clear deviations from Zero Trust Architec-ture (ZTA) principles. To address these shortcomings, this thesis proposes a central-ized Security Gateway positioned in front of MCP servers, enforcing defense in depththrough three layers: malicious input filtering (WAF), request rate limiting, and a Security Token Service (STS) implementing OAuth 2.0 Token Exchange (RFC 8693) toreplace implicit token forwarding with dynamic access control based on the principleof least privilege. This architecture establishes clear trust boundaries and enables acontrolled, scalable, and traceable AI integration in enterprise environments.