Cipher CadetBeginner21 daysFree

AI Red Teaming From Zero: A 21-Day Hands-On Course

In 21 short days you go from never having tested a security system to running a structured attack assessment on an AI application and writing it up.

Taught by Mr Elite 2 of 21 days live A new day every day

Start Day 1 See the syllabus

0of 21 days done

I wrote this course for people who have never done security work. Each day is one lesson of 45 to 75 minutes. I explain what we are doing and why, we do it in a SecurityElites browser lab, and you finish with something I can check: a screenshot, a note or a short write-up. Every attack in this course runs against a practice lab or your own notes, never against a real system.

We start with how AI applications are built, because you cannot attack what you cannot picture. Then we move through prompt injection, jailbreaking, retrieval systems (RAG) and agents that use tools, and we tie each one to the OWASP LLM Top 10. Each module ends with a day where you apply what you learned and review earlier material. On day 21 you run a full assessment and write your first red team report.

By the end you can

  • Explain how an AI application is put together and where its trust boundaries sit.
  • Describe each item in the OWASP LLM Top 10 in plain words with an example.
  • Carry out direct and indirect prompt injection against practice labs and say why each worked.
  • Test a model with common jailbreak techniques and judge how well its guardrails hold.
  • Attack and then defend a RAG system and a tool-using agent in practice labs.
  • Recommend practical fixes such as least privilege, output handling and human approval.
  • Write a clear red team report with evidence, impact and remediation.

Who this is for

  • Complete beginners to security who want a practical way into AI security.
  • Developers, testers and analysts who use or build chatbots and want to know how they get attacked.
  • Career changers who want a portfolio of written lab work.
  • Anyone who can use a computer and a browser and is willing to follow steps carefully.

Before you start

  • Nothing in security is needed. You should be comfortable using a browser, copying and pasting text, and saving screenshots.
  • A free SecurityElites account and about an hour a day.
  • A modern web browser
  • A free SecurityElites account
  • A notes app or text editor
  • A screenshot tool (built into your operating system is fine)

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 AI Apps Work and Break

Understand how an AI application is built and perform your first prompt injection.

  1. 1Orientation and your first winBy the end of today you can set up your workspace and make a practice chatbot ignore its instructions.Lab: AI Prompt Injection
  2. 2How an AI application is put togetherBy the end of today you can draw a chatbot application and label the system prompt, user input, context and output.
  3. 3Threat thinking and the OWASP LLM Top 10By the end of today you can name each OWASP LLM Top 10 item and explain it in one plain sentence.Unlocks 16 Nov
  4. 4Direct prompt injection and system prompt leaksBy the end of today you can extract a hidden system prompt from a practice bot and explain why it matters.Review | Lab: AI System Prompt Extraction | Unlocks 17 Nov

Module 2 Jailbreaking and Guardrails

Learn the main jailbreak families and how to judge whether a guardrail holds.

  1. 5Jailbreaking and the roleplay bypassBy the end of today you can explain the difference between jailbreaking and prompt injection and run a roleplay jailbreak in a lab.Practice | Lab: AI Jailbreak | Unlocks 18 Nov
  2. 6Instruction conflicts and refusal suppressionBy the end of today you can use conflicting rules and prefix injection to test whether a model keeps its refusal.Practice | Lab: AI Instruction Conflict | Lab: AI Refusal Suppression | Unlocks 19 Nov
  3. 7Encoding and language tricksBy the end of today you can use encoded text and a low-resource language to test for gaps in a model's filters.Practice | Lab: AI Jailbreak | Lab: AI Low-Resource Language Jailbreak | Unlocks 20 Nov
  4. 8Multi-turn and many-shot jailbreaksBy the end of today you can use gradual escalation and fake dialogue priming against a practice model and describe the pattern.Practice | Lab: AI Gradual Escalation | Lab: AI Many-Shot Jailbreak | Unlocks 21 Nov
  5. 9Output filters, defences and a jailbreak test planBy the end of today you can bypass a simple output filter in a lab and write a repeatable jailbreak test plan.Review | Lab: AI Output Filter Bypass | Unlocks 22 Nov

Module 3 Indirect Injection and RAG

Attack the data an AI reads, including documents, images and knowledge bases.

  1. 10Indirect prompt injectionBy the end of today you can hide an instruction in a document and show an AI summariser obeying it.Practice | Lab: AI Indirect Prompt Injection | Unlocks 23 Nov
  2. 11RAG explained and knowledge base poisoningBy the end of today you can describe how retrieval-augmented generation works and poison a practice knowledge base.Practice | Lab: AI RAG Poisoning | Unlocks 24 Nov
  3. 12Hidden text: images and invisible charactersBy the end of today you can show instructions hidden in an image and in invisible Unicode characters, and explain why humans miss them.Practice | Lab: AI Visual Prompt Injection | Lab: AI Hidden Character Injection | Unlocks 25 Nov
  4. 13Attacking and defending a document pipelineBy the end of today you can defeat a document summariser with a crafted file and propose three fixes.Review | Lab: AI Adversarial Document Upload | Unlocks 26 Nov

Module 4 Agents, Tools and Memory

Show how tool-using agents are hijacked and how least privilege limits the damage.

  1. 14Agents, tools and excessive agencyBy the end of today you can explain what an AI agent is and judge whether its permissions follow least privilege.Practice | Lab: Coding Agent Blast Radius | Unlocks 27 Nov
  2. 15Hijacking an agent through a poisoned issueBy the end of today you can show how hostile text in a ticket steers a coding agent into an unwanted action.Practice | Lab: Poisoned GitHub Issue | Unlocks 28 Nov
  3. 16Tool descriptions and rogue MCP serversBy the end of today you can review a tool's description for hostile instructions and show a poisoned tool hijacking an agent.Practice | Lab: Rogue MCP Server | Lab: AI Tool Poisoning | Unlocks 29 Nov
  4. 17Memory poisoning and data exfiltrationBy the end of today you can show a persistent injection through agent memory and trace how data leaves through a tool.Practice | Lab: AI Memory Poisoning | Lab: AI Data Exfiltration | Unlocks 30 Nov
  5. 18Agent hijack: apply and reviewBy the end of today you can hijack an agent with an indirect tool call injection and explain the fix.Review | Lab: AI Agent Hijacking | Unlocks 1 Dec

Module 5 Defence, Reporting and Capstone

Turn attack skills into fixes and a professional report.

  1. 19Unsafe output handling and fake packagesBy the end of today you can show how model output becomes an attack when an app trusts it, and spot a hallucinated package name.Practice | Lab: AI Markdown-Rendering Injection | Lab: AI Package Hallucination | Unlocks 2 Dec
  2. 20Defence, remaining Top 10 items and report writingBy the end of today you can map defences to every OWASP LLM Top 10 item and structure a finding for a report.Practice | Lab: AI Prompt Firewall | Unlocks 3 Dec
  3. 21Capstone: assess an agent and plan your next stepsBy the end of today you can run a structured assessment of a practice agent and deliver a report with a learning plan.Capstone | Lab: AI Tool-Output Injection | Unlocks 4 Dec

The final project

On day 21 you run a structured assessment against an agent-based practice lab, using techniques from across the course and mapping each finding to the OWASP LLM Top 10. You deliver a short red team report with evidence, impact and fixes, and then plan what to learn next.

Questions people ask

Do I need any security or coding experience?

No. I explain every term the first time it appears, and each step is written out in full. If you can use a browser and take screenshots, you can follow along.

Is it legal to try these attacks?

Yes, because every hands-on step runs in a SecurityElites practice lab or on your own notes. Never point these techniques at a real system you do not own or have written permission to test.

How much time does each day take?

Plan for 45 to 75 minutes, including the lab and the mission. Some labs in the later modules are harder, so take a break and come back rather than skipping the mission.

Will this course make me a professional AI red teamer?

It gives you a solid beginner foundation, a set of written lab results and a clear route onward. Professional work needs further practice, which is why day 21 ends with your own plan for what to learn next.