Shell WalkerIntermediate14 daysFree
Securing AI Agents and MCP: Hijacking, Poisoning and Lockdown
In 14 days you will learn to break AI agents in a safe lab, then review and lock down the MCP servers and tools they depend on.
Taught by Mr Elite 14 of 14 days live Complete
I built this course for people who are already comfortable in a terminal and want to understand why AI agents are such a rich target. An agent is a language model with tools: it reads files, calls APIs, runs code and talks to other agents. Every one of those abilities is a place where an attacker can step in. Over 14 days I will show you how agent hijacking, tool poisoning, confused deputy attacks and excessive agency actually work, so you can spot them and shut them down.
Each day is one lesson of about 45 to 75 minutes, including practice. I explain the idea, we do it together, and then you complete a mission that produces something I can check: a screenshot, a note, a working command or a short write-up. Everything runs on your own machine or in the named SecurityElites practice lab, never against real third-party systems. The course is split into modules, each ending with a day where you apply what you learned, and it finishes with a capstone where you review an MCP setup from start to finish.
By the end you can
- Explain how an agent, its tools and MCP fit together, and map the trust boundaries between them.
- Demonstrate indirect prompt injection and agent hijacking in a lab, and describe the defences that stop them.
- Recognise tool poisoning and tool output injection, and review an MCP tool definition for hostile content.
- Show how a confused deputy attack misuses an agent's privileges, and design checks that prevent it.
- Identify excessive agency and apply least-privilege permissions to an agent.
- Review an MCP server before connecting it and write a lockdown checklist.
- Produce a written security review of an agent and MCP setup with prioritised fixes.
Who this is for
- Security learners who are comfortable in a terminal and understand how the web and networks work.
- People who have tried a few labs and want to move into AI and agent security.
- Developers and engineers who are starting to connect agents and MCP servers to real systems and want to do it safely.
- Anyone who wants a practical, defensive view of how agents fail.
Before you start
- Comfort with the command line: moving around, running commands, reading output.
- A working grasp of HTTP, JSON and basic networking.
- Some prior lab experience, even a handful of beginner web labs.
- No machine learning knowledge is needed. I explain every AI term as it appears.
- A modern web browser for the SecurityElites labs
- A free SecurityElites account
- A terminal (Linux, macOS, or Windows with WSL)
- Python 3.10 or newer
- Node.js 18 or newer with npm
- A text editor such as VS Code
- curl and jq
Syllabus
One day at a time. Finished days get a tick. Days that are not out yet show when they unlock.
Module 1 Foundations: How Agents and MCP Work
Understand agent architecture, MCP and where the trust boundaries sit, and see a first injection.
- 1Orientation: What an Agent Is and Why It Is RiskyBy the end of today you can describe an AI agent's parts and complete your first prompt injection in the lab.Lab: AI Prompt Injection
- 2MCP Explained: Servers, Tools and TrustBy the end of today you can explain how an MCP client talks to a server and list the tools a server exposes.
- 3Mapping an Agent's Attack SurfaceBy the end of today you can draw a threat map of an agent showing every input, tool and privilege.Practice | Lab: AI Indirect Prompt Injection
Module 2 Agent Hijacking
Hijack agents through untrusted content and tool results, and learn the defences.
- 4Agent Hijacking Through Indirect Tool CallsBy the end of today you can hijack a lab agent into making a tool call the user never asked for.Lab: AI Agent Hijacking
- 5Hijacking a Coding Agent via Poisoned ContentBy the end of today you can show how an issue or document steers a coding agent, and name the controls that would stop it.Practice | Lab: Poisoned GitHub Issue | Lab: AI Tool-Output Injection
- 6Persistence and Exfiltration: Memory, Plans and Data LeaksBy the end of today you can show how an injection persists in agent memory or leaks data through a tool.Review | Lab: AI Memory Poisoning | Lab: AI Data Exfiltration
Module 3 Tool Poisoning and Tool Abuse
Spot hostile tool descriptions and show how tool arguments turn into real damage.
- 7Tool Poisoning: Hostile MCP Tool DescriptionsBy the end of today you can show how a malicious tool description hijacks an agent and find the hostile text in a definition.Lab: AI Tool Poisoning
- 8Rogue Servers, Spoofed Agents and Tool ConfusionBy the end of today you can tell a rogue or spoofed tool provider from a legitimate one and explain how tool confusion gives the wrong effect.Practice | Lab: Rogue MCP Server | Lab: AI Tool Confusion
- 9When Tool Arguments Become AttacksBy the end of today you can demonstrate an injection through a tool argument and describe how to validate arguments safely.Review | Lab: AI Tool Argument Injection | Lab: AI Plugin Path Traversal
Module 4 Confused Deputy and Excessive Agency
Show how agents misuse their own privileges and cut their power down with least privilege.
- 10Confused Deputy: Turning the Agent's Privileges on Its UserBy the end of today you can show a confused deputy attack in the lab and explain which identity check was missing.Lab: AI Confused Deputy
- 11Excessive Agency and Least PrivilegeBy the end of today you can reduce a coding agent's permissions to the minimum its task needs.Practice | Lab: Coding Agent Blast Radius
- 12Human Approval: Where It Works and Where It FailsBy the end of today you can show how an approval prompt can be spoofed and design one that is harder to fool.Review | Lab: AI Human-in-the-Loop Bypass | Lab: AI Agent TOCTOU Race
Module 5 Lockdown and Capstone
Review and lock down an MCP setup, then prove it all in a full review.
- 13Reviewing and Locking Down an MCP ServerBy the end of today you can run a structured review of an MCP server and produce a lockdown checklist.Practice | Lab: Rogue MCP Server
- 14Capstone: Security Review of an Agent and MCP Setup, and What NextBy the end of today you can deliver a prioritised security review of an agent and MCP setup, backed by proof of at least two findings.Capstone | Lab: AI Plan Injection | Lab: AI Recursive Self-Prompting
The final project
You will review a deliberately flawed agent and MCP server setup in the practice labs, find the hijacking, poisoning, confused deputy and excessive agency problems, and prove at least two of them. Then you will write a short security review with a lockdown plan, ranked by risk.
Questions people ask
Do I need to know machine learning to take this course?
No. You need terminal confidence and a grasp of web basics. I explain every AI and MCP term the first time it appears, and we only cover as much model behaviour as the attacks require.
Is it safe and legal to practise these attacks?
Yes, as long as you stay in the course setup. Every attack runs against your own machine, a local test server or the named SecurityElites lab. Never point these techniques at systems you do not own or have permission to test.
How much time does each day take?
Plan on 45 to 75 minutes, including the mission. Some labs are marked advanced, so on those days it is fine to use the full time or slightly more, and to repeat the lab if it does not click first time.
Will this help me defend real agents and not just attack labs?
That is the point. Each attack day ends with the controls that stop it, and the last two days have you review and lock down an MCP setup and write up the findings as you would for a real team.