Застосування Zero-Trust Architecture для забезпечення безпеки AI-агентів на базі Model Context Protocol

Abstract

This paper proposes an approach to integrating Zero-Trust Architecture (ZTA) principles, as defined by NIST SP 800-207, into the security layer of AI agents operating via the Model Context Protocol (MCP). The MCP specification enables LLM-based agents to invoke external tools through a standardized JSON-RPC 2.0 interface, but does not prescribe a comprehensive security policy. The proposed method adapts the Policy Engine, Policy Administrator, and Policy Enforcement Point triad to intercept and validate every MCP tool invocation. Microsegmentation is applied at the MCP server level, ensuring each server operates within an isolated trust zone with least-privilege access. Continuous authentication monitors agent behavior patterns to detect anomalous action sequences, enabling real-time revocation of permissions.