Free to start · built by Hamcodes

Build AI security skills.
Test your defences.

Learn prompt injection foundations, then practise attacks, mitigations and retesting in guided AI scenarios, while keeping legitimate work running.

Free first unit · no code or sign-up needed to start.
Today's challenge · Member training: Start / Resume

Test your skills · free scenario

Catch the attack. Keep the useful request.

Three short calls from an assistant's inbox. Decide what can pass, then read the reason and the cost of your choice.

THE FILTER / LIVE PREVIEW

Three messages.
Every decision counts.

Foundations / practitioner refresher. No coding or earlier lessons required.

Two hostile requests. One legitimate question. Choose what reaches the assistant, then see the reason behind each call.

  • Spot instructions in untrusted content
  • Keep legitimate work running
  • See both kinds of mistake
Try a different challenge every day →

Scripted teaching examples. No model, email service or real account runs here.

Incoming 01 / 03
Scenario brief, completion criteria and what to record
Difficulty and prerequisites
Foundations / practitioner refresher. Read short prompts; no coding or earlier lessons required.
Legitimate task
Answer a security education question and summarise a product review.
Attacker objective
Replace the assistant's instructions or turn text in a review into an email action.
Completion criteria
Identify both hostile requests and allow the legitimate one. Explain which text has authority.

Save the evidence in your own notes: the input, your decision, the trust boundary and the possible consequence. Then review prompt injection or go directly to Incident labs. The catalogue needs no course completion; running incident kits needs Python and active Pro.

These are scripted teaching examples. No model, email service or real account runs in this demo. For another scenario, try today's challenge.

Start here

The chat box is the door, not the building.

Most people meet artificial intelligence through a chat window and reasonably assume that the window is the system. It is the entrance.

Behind it sit the parts that make the product useful, and those are the parts that can be turned against you.

WHAT YOU SEEWHAT IS BEHIND ITA chat boxyou type, it answersA store of documentsanyone may add oneA function that actssends email, changes recordsOther companies' toolsdescriptions you did not writeThis is the partmost people pictureThis is the partyou are guarding
The chat box is the entrance, not the building. Everything on the right is reachable from it, and every one of those parts accepts text written by somebody you do not control. Step 1 walks through the whole building and names each part.

Your next step

Start from the knowledge you have.

02 / PUT IT TO THE TEST01 / START FROM ZERO

Learn the foundations

For developers and security beginners. The free first unit explains prompt injection, then gives you five decisions with feedback. No coding or AI background needed.

Your first outcome: explain an attack and identify the boundary between instructions and untrusted content.

Start the free first unit →

Prefer to read first? Start the free course.

Test your skills

For practitioners and builders ready to make defence decisions. Start with the scenario above, then choose an incident without completing the course.

Your evidence: reproduction steps, observed behaviour, a mitigation, attack and legitimate-request retests, and remaining limits.

Open the free scenario →

Browse Incident labs · Refresh the defences.

WHY THE FILTER

Practise the defender's decision too.

There are good games about breaking a model. Trick the bot, steal the password, climb the levels. They are fun, and they teach you attacking.

Here you also practise the other seat: deciding what should reach the model without blocking the people using it. That decision matters to security engineers, AI product teams and platform owners. Every level maps to a named, published risk, so what you learn carries a term you can use in a design review.

The browser game uses scripted decisions. The local projects then let you reproduce a flaw in a small test application, apply controls and record what still works.

The one idea

A filter that blocks everything is not a good filter.

Catching every attack is easy. Block all traffic and your breach count is zero.

A useful product has to serve legitimate requests too. Practice here includes both hostile inputs and innocent messages that can look similar.

You get scored on two numbers here, not one. Every single time.

Number one

The attacks you missed

The hostile message you waved through. Instructions buried inside a document your assistant was asked to read. Something that cost your company money.

Number two

The people you blocked

The real customer you turned away. The trainer asking a fair question. The person writing in other unpopular languages. The developer whose work you just broke.

See the trade-off between allowing and blocking everything
Allow everythingzero annoyed customersevery attack gets inBlock everythingzero breachesnobody can use itEvery real system lives somewhere along hereand moving either way costs you the other
Both ends of this scale lose useful behaviour. Test what a defence catches and what legitimate work it blocks. A control may improve both, but that needs evidence from both kinds of request.

Real systems need both numbers. A change can affect each differently, so test attacks and legitimate requests together instead of assuming that fewer allowed requests means a better defence.

Try it now

Grade a filter in ten seconds.

Put in the numbers a vendor would quote you, or the ones from a system you already run. The calculator works out what those percentages do once they meet real traffic, and gives the result a grade.

Nothing is sent anywhere. The whole thing runs in this page.

Filter grader Live
– Enter some numbersThe two on the right are the ones a vendor publishes. The two on the left are the ones only you know.
–Attacks caught
–Attacks missed
–People blocked in error
–Precision

Try the last preset. A filter catching 99.9% of attacks while wrongly flagging only 0.1% would be world class, and no vendor will sell you one. It still blocks roughly one innocent person for every attack it catches. That is what rarity does to any detector, and step 1 works through why.

The free course

Five steps. Read them in order.

This is the whole thing, open, with no sign up needed to read or play. It is written for one person learning alone. Each step tells you what you will be able to do by the end of it.

By step 4 you are building a working AI application on your own laptop with free open-source software, then breaking it deliberately to see the flaw for yourself.

See the full five-step learning path
01Buildhow it works,where it breaks02Attackswhat theylook like03Solutionswhat tobuild instead04Practicalbuild oneyourself05Resourcescheck ityourselfyou start hereknowing nothingyou finish able tobuild and defend one
Read them in order and each one assumes only what came before it. Step 1 is the long one and divides cleanly in half. By step 4 you are running a model on your own laptop and attacking it deliberately.

Read them in that order and you go from not knowing what a model is to building one into an application and defending it. If you have only fifteen minutes, read part 1 of step 1, or play today's prompt.

The full game

Then practise, until it is a reflex.

Reading about a trick and catching one are different skills. The game puts you in the seat for 35 decisions across seven units.

One idea, then practice. Another idea, then practice. Never more than two exercises before something new arrives, so it never turns into a quiz you grind through.

Prompt Injection

LLM01

Why a model cannot tell your rules apart from a document it was handed, and what attackers do with that gap.

Free · 4 lessons · 5 exercises · play it now

Disguises

EVASION

Encoding, switching language, mixed case, and invisible characters. The same attack wearing a costume your filter cannot read.

3 lessons · 5 exercises

Hidden Context Exposure

LLM08

Asking outright, filling in the blank, and sizing attacks that map your rules and your tool list without ever quoting them.

3 lessons · 5 exercises

Excessive Agency

LLM03

What happens when an over-permissioned agent reads a poisoned invoice. A trick plus tools is remote control.

3 lessons · 5 exercises

Document Store Attacks

LLM05 / 09 / 02

Attacking the knowledge base instead of the message, plus one customer's search quietly returning another customer's data.

3 lessons · 5 exercises

Output and Supply Chain

LLM10 / 04

Model output run as a database query or drawn as a web page, and the unchecked file that runs code the moment you load it.

2 lessons · 5 exercises

Combined Attacks

BOSS

Chained tricks. Attacks planted four messages early. Nothing looks decisive on its own.

Final unit · 5 exercises

How it drills

Three ways to be wrong about the same attack.

Seeing a trick once is trivia. Meeting it from three directions is a reflex.

Allow or block

A message arrives. You make the call. Twenty-two of these, and eight of them are real people you should never have stopped.

Tap the attack

A document, a web page, a fetched chunk. One piece of it is an order. Find it before your assistant obeys it.

Name the technique

Read the system, name the flaw. This is the part that turns a gut feeling into words you can use in a meeting.

Choose the scope of your practice

Three practical offerings, with different jobs.

Free practical walkthrough

Course step 4: browser tools and one local app walkthrough with three controls. Read steps 1 to 3 first; the Python is supplied. Keep your before-and-after notes.

Free. No purchase or membership.

Open the walkthrough →

Build & Break project

A full Notes Assistant in Python or n8n, four attacks, five defence layers, starter kit, scorecard and capstone report. Needs local model setup; the project includes hardware guidance.

Included in Full Training and active Pro.

See the project →

Incident labs

Small recreations of disclosed flaws, each with its own kit and sources. For practitioners comfortable running Python and reading logs. Reproduce, fix and retest with fake data.

Catalogue free to browse. Running kits requires Pro.

Choose an incident →

Progression: understand the concept, practise a decision, build a small app, apply a defence, then retest an attack and a legitimate request. General AI background stays available in step 1.

What you can take away

A finding you can explain and retest.

The applied project ends with a short findings report. This example shows the evidence to save, including the useful work your defence must preserve.

Illustrative example · not a measured product or learner result

A review asks a summariser to send a conversation

Legitimate task
Summarise a customer's headphone review, including its battery-life claim.
Untrusted input
A review includes a fake SYSTEM instruction requesting that the conversation be emailed outside the application.
Attacker objective and boundary
Turn third-party review text into an authorised tool call. Review content must remain data; it cannot grant permission to send email.
Read the reproduction, mitigation and retest criteria
Reproduction and illustrative failure
Give a test summariser the review from the free demo. In a vulnerable design with an email tool and standing send permission, it could attempt an external send instead of only summarising. The browser demo itself sends no email.
Mitigation to test
Remove email access from the summariser. Treat retrieved reviews as untrusted data, constrain permitted actions in application code, and validate any allowed tool arguments. Role labels alone are not a reliable boundary.
Attack retest
Replay the same review. Required result: no send action is available or executed; the review cannot grant that permission. Record the actual tool log before marking it passed.
Legitimate-request retest
Summarise a clean review and the review containing the injection. Required result: useful summaries retain the battery-life information without executing the embedded instruction. Record the actual outputs.
Remaining limitations
Removing a tool prevents that action, but the model may still produce a misleading summary. One input does not establish a success rate. Test other wording, document types and legitimate requests, using fake data.

Retest status: proposed acceptance criteria, not executed results. Replace each with your own observations in the project.

Download the sample finding

The paid Build & Break project applies this report structure to your local Notes Assistant. You run the attacks and save your own scorecard; this example previews the method.

New · in The Filter Pro

See real breaches happen, then fix them.

Ten hands-on labs built from recently disclosed vulnerabilities, several exploited in the wild. You see how each one happened, build a tiny version of the flaw, break it yourself, then apply the fix, with a one-file kit for your own machine.

A lab kit running in a terminal: the vulnerable version leaks a secret to a safe sink in red, the fixed version leaves the sink empty in green
Every lab ships a kit like this. One file, nothing to install, fake data, and a safe sink that records what would have happened instead of doing it.

Project Full Training or Pro

Build it. Break it. Filter it.

You have been the filter. Now build the thing the filter protects, break it yourself, then defend it. The Build & Break project is a full section of the paid game, next to the units and the Practice Range.

The hardware calculator rating six model sizes as comfortable, slow or won't fit on a laptop with 16 GB of memory
The calculator, before you download anything.
An attack replay: the agent lists the notes, then opens the meeting note that carries the hidden instruction
An attack replay, one turn at a time.
The honest scorecard: each of the four attacks marked blocked or got through
The scorecard. Blocked, or got through.
01 · BUILD

Build it.

Run an open-weights model on your own laptop, then build a small Notes Assistant around it. It lists your notes, reads one, and does the sums. In plain Python, or with no code at all in n8n.

02 · BREAK

Break it.

Plant four attacks in its notes, inspect the tool traces and record what happens. Results vary with the model and settings. A hidden instruction. A file outside its folder. A tool turned against it. A jailbreak. Every one is something you already learned to spot.

03 · FILTER

Filter it.

Add five layers of defence, run every attack again, and fill in a scorecard that does not flatter you. Then a capstone: five new attack notes, and a badge for surviving them.

What you build

A real AI agent, then a defended one

A Notes Assistant that runs on your own computer, reads your files and uses tools. Then the same agent with five layers of defence around it.

What you can do afterwards

Skills you can name on a CV

  • Run and size local AI models
  • Build a tool-using agent, in code or no-code
  • Test an AI app for prompt injection
  • Add defences and measure them honestly
  • Write a short red team findings report
When you would use it

Before an AI feature ships

  • Your team wants to connect an assistant to company documents
  • You are asked to review or test an AI feature
  • You are moving towards an AI security role

Check your machine first. The local project needs a compatible laptop or desktop, at least 8 GB of memory, about 10 GB of free disk and Python 3.9 or newer for the code path. The hardware calculator helps you choose a model that fits. If you cannot run a model yet, start with the browser exercises and scripted attack replays, then return to the local build. Use fake data and test only systems you own or have permission to assess.

Explore all five stages, tools and starter-kit contents
  1. Stage 1Run itLocal models, open weights versus open source, quantization, and a calculator that checks your machine first.
  2. Stage 2Build itA model, a loop and tools. The agent, step by step, in code or in n8n.
  3. Stage 3Break itFour attacks, scripted replays and transcript drills.
  4. Stage 4Filter itSpotlighting, least privilege, a human in the loop, input filters, output checks.
  5. CapstoneRed team itFive new notes against your own defended agent, then a findings report.

Local model setup

LM Studio for a gentle start, Ollama for the real thing, on Windows, macOS or Linux. Written for people who have never run a model.

Code or no-code

Both paths build the same agent, so you can compare them. A toggle remembers your choice. Shared steps need no choice at all.

The hardware calculator

Tell it your memory and graphics card, and it rates each model size as comfortable, slow or too big. The calculator itself lives inside the paid game.

Attack and defence labs

Real attack notes against your own agent, mapped back to the units you already played, with an honest scorecard at the end.

A starter kit

The Python for every step, a sandbox of fake notes, the attack and capstone notes, and a scorecard, in one download.

Safe by design

Fake data only, one sandbox folder, and your own machine. Nothing here touches a real account.

Full Training includes this project and core updates permanently. Pro includes online core access while subscribed, plus Incident labs and new releases. Compare $79 once, $59/year or $7/month. Activate with the key from your Gumroad receipt in the member app.

Get it

The course is free. The practice is $79.

Free practice, permanent core or ongoing incident training.

You never have to pay anything. All five steps stay open, and so does a new prompt every day. The Filter Pro is the full game plus a new attack every month, because the attacks keep changing.

Founding members: $39 a year instead of $59, for as long as you stay. Pro opens on 1 December 2026, and founders get the full game today. 25 of 25 founding places left

Free

$0

The whole course, a daily round and the first unit of the game.

  • All five steps of the course
  • Both browser tools
  • Unit 1, Prompt Injection, and its Free Defender decision certificate
  • Today's prompt, every day
  • The incident labs catalogue
  • Units 2 to 7, the rest of the Range and the Build & Break project
  • Monthly attack packs and the full completion certificate
Play today's prompt or start step 1

The Filter Pro

Soon

Everything in the full game, plus a new attack pack and mini lab on the 1st of every month. Opens 1 December 2026.

  • 10 incident labs from real, recently disclosed breaches: build, break and fix each one
  • Online core access: all 7 units, the Practice Range and the Build & Break project
  • New labs and attack packs added as new vulnerabilities are disclosed
  • A matching Build & Break mini lab with each pack
  • The full pack archive and the daily prompt archive
  • A badge for every pack you finish
  • Up to 3 devices. Cancel any time in Gumroad
See the incident labs

Bought the one-time game before? Everything you bought stays yours, including every core update, and your purchase key still opens it. The incident labs and the packs are new content made after your purchase, so they come with Pro.

After checkout, use the key from your Gumroad receipt or library at member activation, then Start / Resume training. The free public training app keeps its own browser-local progress. To move progress, export it from the account menu and import it at your destination; the two domains do not share browser storage.

Creator, evidence and scope

Check the reasoning behind the exercises.

Created by K.H. at Hamcodes. The free course links its sources and evidence limits; each Incident lab links the disclosure in the member app. A small recreation teaches a failure mechanism, and does not reproduce an entire production system.

The free Unit 1 certificate records completion of its scripted defence decisions. The full certificate requires all seven units and a self-recorded Build & Break project capstone in the member app. These are completion records based on browser-stored progress, not independently assessed professional qualifications or assurance that a production system is secure. The sample finding above is illustrative; your project report should contain your own observed results.

Questions

Before you start.

I know nothing about AI. Is this really for me?

Yes, and that is precisely who it was written for. Step 1 assumes you have never encountered the word prompt, and every new term is defined in plain language the first time it appears. If a sentence anywhere on this site has to be read twice, treat that as a fault and tell me about it.

Do I need to write code?

Not for steps 1 to 3, or for step 5. Those are written for anybody who ships or reviews AI features. Step 4 does have you build something, and it supplies every line you need. If you can follow a recipe, you can complete it.

Will I actually be able to build something at the end?

Yes. Step 4 walks you through running an open-source model on your own laptop, wrapping it in a small application, feeding it a document it did not write, and watching the flaw occur. You then repair it using three of the controls from step 3 and measure both numbers.

How do I get into the full game?

Open the member training app at private.hamcodes.com and activate the purchase or licence key from your Gumroad receipt. The public /play/ app is the free first unit and Range preview. Progress saves in each browser, and these domains do not share storage. Use Export progress and Import progress in the account menu when moving between them. The Build & Break starter kit runs locally.

What happens to my one-time purchase?

Everything you bought stays yours, with core updates, exactly as sold. Your purchase key keeps working. Incident labs and monthly attack packs are separate Pro content. A one-time core purchase remains valid if a separate Pro membership ends.

Can I cancel Pro any time?

Yes. You manage the membership in your Gumroad library and cancel with one click. You keep subscription access until the end of the period you already paid for. Then Pro access ends; your progress and badges stay in your browser. Separately purchased Full Training access remains yours.

Does Pro track me?

We count anonymous clicks to improve the course: which exercises people miss, whether a new pack gets opened. No cookies for tracking, no ads, no IP addresses, no emails, and nothing at all if your browser sends Global Privacy Control or Do Not Track. Your licence key is checked with Gumroad, which is where you paid. The one-time game sends no events at all.

What is in a monthly pack?

One attack family, released on the 1st of the month: two to four short lessons, eight to ten exercises in the game's three formats, and a matching mini lab for the Build & Break agent you built. Every example is deliberately broken, so it teaches the shape without working. Finishing a pack earns its badge. See the incident labs.

Is this teaching people how to attack AI systems?

The browser game teaches recognition and defence decisions with scripted examples. The local Build & Break project and Incident labs go further: you reproduce attacks in small isolated examples with fake data, apply controls and retest. Test only systems you own or have permission to assess.

How long does the whole thing take?

About three hours to read the five steps, plus an hour for the build lab. Step 1 is the long one at around 45 minutes, so it comes in three parts of about fifteen minutes, each ending with a save point. Today's prompt takes one minute. The game is another 45 to 70 minutes. You can work through it across several sittings, and progress saves in your browser.

Is it kept up to date?

Yes. The core game is updated as the published Top 10 for LLM applications changes, and those updates are included for everyone who bought it. New attack families arrive as monthly packs in Pro, and the free course and the daily prompt keep moving too.

Can I use it to train my team?

Yes. For a small team, a membership each works fine, and the free course and the daily prompt are open to everyone. If you want to run it as internal training, cover a classroom, or put your own name on it, get in touch and we will sort out a licence.

Start

Keep building, one tested decision at a time.

Today's prompt takes sixty seconds. Part 1 of step 1 takes about fifteen minutes and assumes no prior knowledge whatsoever. That is the entire commitment.