
Introduction to the course.
Understand how the Model Context Protocol (MCP) enables AI to interact with tools, data, and actions through dynamic capabilities and schemas, and why discovery raises security risks and attackers’ opportunities.
Clarify that MCP is a protocol while the wire is a transport, with HTTP plus SSE and htdio as common transports enabling long-lived, streaming sessions.
Discover the MCP pentesting workflow from initialization to enumeration of tools, resources, and templates, and learn how missing built-in authorization makes enumeration the attack surface.
Install the red team MCP client and run the client-generic.py script with uv to connect to an HTTP+SSE server, inspect the MCP protocol with Wireshark, and explore tools and resources.
Explore command execution and file disclosure using safe tools, showing how restrictions can be bypassed and attackers may access sensitive system files.
Explore challenge 3 — excessive permission scope by using the file manager tool to read /etc/passwd and /etc/shadow, and learn to restrict permissions and sanitize input to prevent abuse.
This lecture explains tool shadowing on MCP servers, showing how a malicious server could imitate a trusted tool, and why single-server setups may limit exploitation opportunities.
Demonstrates authenticating with username and password and token verification, revealing a flawed verify token implementation that allows logging in as any user despite incorrect tokens.
Explore token theft and validation failures by evaluating expressions and generating code examples, using a JSON input to execute OS commands and exposing system control risks.
AI systems are no longer passive chatbots. Modern AI agents read files, invoke tools, execute code, and interact directly with internal systems and infrastructure. This shift fundamentally changes the security model — and it introduces an attack surface most organizations are not prepared to defend.
The Model Context Protocol (MCP) is rapidly emerging as the standard interface that allows AI agents to discover tools, access resources, and perform actions. While MCP unlocks powerful capabilities, it also creates new trust boundaries, implicit permissions, and protocol-level weaknesses that traditional application security approaches fail to address.
This hands-on course is designed to close that gap.
You will learn how MCP actually works at the protocol and transport level, how AI agents reason about tools and resources, and where real-world implementations go wrong. From there, you’ll move directly into offensive techniques: enumerating MCP servers, abusing exposed tools, exploiting over-trusted resources, chaining capabilities, and breaking assumptions built into poorly designed MCP deployments.
This is not a prompt engineering course.
This is not about tricking language models with clever text.
This is infrastructure security for AI systems.
You will attack MCP servers the same way real red teams would — by understanding protocol flows, abusing trust relationships, exploiting design flaws, and demonstrating impact. You’ll also learn how to defend MCP environments, identify insecure patterns, and design MCP servers that are resilient against real adversaries.
The course uses a purpose-built vulnerable MCP lab environment so you can safely practice exploitation and defense techniques without shortcuts or simulations.
If you want to understand how AI systems fail in the real world — and how attackers will exploit them — this course is for you.