AI security engineer
Learn to find, exploit in a lab and fix the weaknesses in AI-powered applications, using AI tools in your own security work from week 1, in 20 weeks.
Coming soon · planned, not written yet
- 20
- weeks planned
- ~30 h
- a week
- 5
- milestone projects
- 1
- project, grown every week
These are planned figures. The weeks are written after we check the tools against current job postings and official documentation, so the final route may differ.
Built for the AI era
How AI fits this roadmap
Companies are putting language models and agents into products that read emails, call tools and touch company data, and each of those links is a new way to attack. This roadmap teaches you to use AI assistants to speed up your own testing and scripting, and to check every result by hand. It also makes AI the thing you defend: prompt injection, unsafe tool use, data leaks and abuse of models, tested and fixed in a lab you build.
The main project
GuardRail Lab
A small AI assistant app with tools and documents that you attack in a private lab, then defend, test automatically and write up like a real security assessment.
What you will be able to do
- 01
Build a small LLM-backed app with tool calling in a private lab, and explain how data and instructions flow through it.
- 02
Demonstrate direct and indirect prompt injection against your own app, and fix each finding with a change you can re-test.
- 03
Map every finding to the OWASP Top 10 for LLM Applications and to MITRE ATLAS techniques in a written threat model.
- 04
Run automated red-team scans with garak, PyRIT and promptfoo, and separate real findings from noise by reproducing them by hand.
- 05
Write a security assessment report with evidence, severity and fixes, including a section on what your AI assistant got wrong.
01The route
Five months, one month at a time.
Each month ends in a milestone project, so you always have something to show for the work.
Month 1 · Weeks 1–4
Security basics and AI-assisted work
Web and API security basics, threat modelling and a safe lab setup. You set up an AI coding assistant, learn to prompt it for scripts, and test what it writes before you trust it.
MilestoneA working private lab with a small web API, a threat model for it and a log of AI suggestions you accepted and rejected
Month 2 · Weeks 5–8
How LLM apps work and how they break
Build an LLM app with documents and tool calling. Learn the OWASP Top 10 for LLM Applications and attack your own app with direct and indirect prompt injection.
MilestoneGuardRail Lab running, with a written list of reproducible attacks against it
Month 3 · Weeks 9–12
Defences for prompts, tools and data
Input and output checks, least-privilege tools, human approval for risky actions and protection of private data. You re-run every attack to prove each fix works, and learn where each fix still fails.
MilestoneA hardened GuardRail Lab with before and after results for every attack
Month 4 · Weeks 13–16
Automated red teaming and agents
Scan with garak, PyRIT and promptfoo, then test agents that browse, read files and call other tools. You use AI to widen your test cases and confirm each finding yourself.
MilestoneA repeatable test suite that runs against GuardRail Lab on every change
Month 5 · Weeks 17–20
Threat modelling, reporting and job search
Map findings to MITRE ATLAS, write the assessment report, plan monitoring for abuse, then prepare a portfolio, interviews and applications.
MilestoneA full security assessment report, a portfolio write-up and a working job-search pipeline
03What is inside
Everything a week comes with.
Every week is built the same way, so once you have done one you know how all of them work.
- 01
One project, GuardRail Lab
You grow one AI app from first build to hardened, tested and reported, so every technique lands in something real.
- 02
A day-by-day plan
Each week has a Learn part and a Do part, with a checklist so you know when you are done. Most weeks include a task to do with an AI assistant and a step to check its work.
- 03
A project in every milestone
Five milestones, each ending in something you can show and link to.
- 04
Reference topics
Short topic pages with check questions you can come back to.
- 05
A job track
Interview practice, portfolio and resume help, and an application plan.
- 06
Progress in your Study plan
Add weeks to your Study plan, give them dates, and finished days are counted for you.
04Is it for you
Who this is, and is not, for.
Better to know before you start than six weeks in.
It is for you if
- You know basic security ideas or can write small programs in Python, and you want to specialise in securing AI systems.
- You are a developer, security analyst or AI engineer who needs to defend AI features at work.
- You can give about 30 hours a week for five months, part-time alongside other commitments.
It is not for you if
- You have never used a terminal or written code. Do a beginner course first.
- You want a job guarantee. Nobody can promise that, and this roadmap does not.
- You want to attack systems you do not own. Every exercise runs in your own lab, on targets you control and are allowed to test.
- You want to trust AI output without checking it. This roadmap trains you to verify it.
?Questions
Before you start
- Is this roadmap ready?
- Not yet. This is the plan. The weeks are being written and checked against current job postings and official documentation. Leave your email and we will write to you when it opens.
- What will it cost?
- The price is not set. Roadmaps are bought once, with no subscription. We will announce it when this one opens.
- Will the plan change?
- It may. Before the weeks are written we check the tools against current job postings, so the lengths and months here are planned, not final.
- Is it legal to practise these attacks?
- Only on systems you own or have written permission to test. Every exercise in this roadmap runs in a lab you build yourself, and it will say so clearly. Never point these tools at someone else's system.
- Will it get me a job?
- It will give you skills and proof of work. It cannot promise a job on a fixed date, because hiring depends on the market, your location and your past work.
Not open yet
Leave your email and we will write once, when it opens.