Unified security agent method and system for large model agent based on zero trust architecture

Abstract

The application discloses a kind of unified security proxy methods and systems of big model intelligent agent based on zero trust architecture. The method comprises the following steps: MCP Server completes trusted access registration and hardware-level remote certification in trusted execution environment, and generates risk portrait through adversarial fuzz testing;In response to user session request, the security gateway uses a special large language model to analyze natural language intent, and dynamically generates a one-time limited token;Extract semantic skeleton before instruction execution and generate semantic fingerprint for integrity check;Real-time construction of dynamic behavior heterogeneous graph within the session cycle, call pre-training graph neural network for cross-tool behavior link anomaly detection;At the same time, the full-link audit log is stored after encryption and anchored by blockchain. The application solves the problem of security protection missing of existing MCP interaction full link, and significantly improves the overall security of big model intelligent agent system.

Images