Paper 2025/1790

CA-MCPQ: A Context-Aware Post-Quantum Protocol for AI Agent Integrity and Security

Abstract

We propose CA-MCPQ, a context-aware post-quantum-secure extension of the Model Context Protocol (MCP). Unlike standard MCP, which leaves authentication, encryption, and authorization optional or implementation-specific, CA-MCPQ elevates them to mandatory protocol-level mechanisms. The design incorporates post-quantum mutual authentication, KEM-derived session keys, and authenticated sequencing to ensure session integrity and prevent replay. Role-based access control is enforced, while token-based authentication is eliminated entirely. AI dynamically infers the required security tier from contextual information and negotiates compatible PQC algorithms; each response returns a reliability score that quantifies alignment with the requested security level. This architecture addresses critical vulnerabilities of MCP—including token misuse, session hijacking, impersonation, and quantum attack—while preserving interoperability. Notably, our evaluation shows that the cryptographic and transport overheads are negligible compared to model computation, confirming that strong post-quantum assurances can be achieved without degrading system performance. Overall, CA-MCPQ provides a practical path toward secure-by-design AI agent ecosystems and lays the groundwork for future extensions such as agent–agent secure coordination.