Securing Agentic AI: A Comprehensive Threat Analysis of Model Context Protocol Systems with Layered Defense Strategies
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
The rise of agentic artificial intelligence, powered by the Model Context Protocol (MCP), creates a complex and evolving security landscape. This article presents a comprehensive analysis of the MCP threat ecosystem through a structured five-layer architectural model: Prompt \& Reasoning, Tool \& Supply Chain, Execution \& Configuration, Protocol \& Network, and Data \& Telemetry. We examine over twenty distinct attack vectors ranging from well-documented prompt injection to novel threats like composability chaining and MCP sampling manipulation, and their manifestations across each architectural layer. Drawing from recent security research by Microsoft, CyberArk, Palo Alto Networks, and others, we analyze real-world vulnerabilities including CVE-2025-32711 (EchoLeak) and CVE-2025-6514 to demonstrate how traditional attack vectors evolve in agentic environments. We present actionable defense strategies, highlighting how modern frameworks like Strands Agents provide essential security controls through identity management, comprehensive observability, and zero-trust tool execution architectures for building secure, observable, and resilient agentic systems.