Set up your computer — what this is, who it is for, and the model every lesson runs on

2 · Pick a tool, against its real numbers

You can skim this table. If you only want to get going, none of it is required — Route B in the next step but one runs a model on your own computer for nothing, with no account and no card. The table is here so that when you do choose a paid tool, you choose it on the two numbers that actually decide it rather than on the advertising.

The prices are per month and the free tiers are what you get without a card. Every name links to its own sign-up page.

Before you sign up for anything — the age rule

Every tool in that table needs an account, and every one of them sets a minimum age in its terms. Broadly: 13 is the floor, 18 is the default, and between the two you need a parent or guardian to set the account up and agree to the terms. The exact rule differs by provider and by country, it is on the sign-up page, and this page is not the authority on it — read theirs.

If you are under 18, the shortest honest path is:

1. Ask a parent, guardian or teacher to make the account. That is what the terms require, and it is a two-minute conversation rather than a problem.
2. Or take Route B below and run a model on your own computer. No account, no card, no age gate, nothing leaving the machine. It is slower and the answers are weaker, and it is a completely legitimate way to do this entire commons.
3. Or do the third way of every lesson, which reads the code and the diff and needs no model whatsoever.

Nothing in the commons asks you to lie about your age, and if a school or club is running this, option 1 or 2 is the answer rather than thirty individual accounts.

3 · Open a terminal and get the files

Four steps here, then one more in section 4. Do them in order. Find your own computer in each step and follow only those lines — the commands really are different between Windows and a Mac, and mixing them up is the most common way this goes wrong.

Step 1 — open the terminal

A terminal is a window where you type commands instead of clicking. Your computer already has one.

On Windows: press the Start button. Type powershell. Click Windows PowerShell.

On a Mac: hold down Command and press the space bar. Type terminal. Press Enter.

On Linux: hold Control and Alt, and press T.

A window opens with a blinking cursor. It is waiting for you to type one line and press Enter. Nothing on this page can damage your computer.

Step 2 — check you have the two programs you need

You need git (it copies the project onto your computer) and Python (it runs the lessons). Type one line, press Enter, read what comes back, then type the next.

On Windows, type these two:

git --version
python --version

On a Mac or Linux, type these two:

git --version
python3 --version

The difference matters. On a Mac the command is python3. On Windows it is python. If you type python3 on Windows you will usually get the Microsoft Store opening, or a blank line. That is not a mistake you made — it is simply the wrong word for that computer. (py -3 --version also works on Windows.)

What a good answer looks like. Something like git version 2.43.0 and Python 3.11.4. Any git 2.x is fine. Python must be 3.10 or newer — if it says 3.9 or lower, install a newer one below.

If it says command not found (or on Windows, is not recognized), you do not have that program yet. Install the one that is missing:

• git: git-scm.com/downloads
• Python: python.org/downloads

Windows readers, one extra thing: in the Python installer, tick the box that says Add python.exe to PATH before you click install. It is switched off by default, and leaving it off is the single most common reason the later commands fail.

After installing, close the terminal window and open a new one, then run the checks again. A terminal only notices a newly installed program when it starts up.

Step 3 — copy the project onto your computer

Two lines, the same on every computer. The first downloads the project. The second moves you inside the folder it just made — like double-clicking a folder to open it.

git clone --branch master https://github.com/spbreed/cyber-commons.git
cd cyber-commons

That is called cloning. You now have every lesson, every skill and every line of the example application on your own computer.

Step 4 — ask your computer whether it is ready

What this is for. You have just typed several commands and read the answers yourself. This one does that checking for you, on the machine you are actually sitting at, and tells you what to fix. Run it now, and run it again any time something later does not work — it is the first thing to try when you hit an error.

On Windows:

python scripts\doctor.py

On a Mac or Linux:

python3 scripts/doctor.py

It prints one line per thing it checked, and each line starts with one of three markers:

• ok — that thing is fine, nothing to do.
• FAIL — something required is missing. The command that fixes it is printed underneath, already written for your operating system, so you can copy it straight into the terminal.
• -- — something optional is missing. You can start the lessons anyway. It tells you what you would gain by adding it.

The last line is the summary. If it says you are ready, you are ready — even if some lines showed --. The third way of doing any lesson needs no AI model at all, so a missing model is never a blocker.

Is there anything to install? For the lessons, no — they use only what comes with Python. The example application you start building in A1.1 does need a few extra pieces, because it is a real web service; A1.2 gives you that one command when you get there.

4 · Tell it which model to ask

This is the one step that matters. You are giving your computer the equivalent of a phone number for a model, so that when a lesson has a job to do it knows who to call.

There is an easy way and a complicated way, and you only need one of them. Take the easy one unless you have a reason not to.

Route A — use a commercial code-generation LLM (the easiest)

Use an AI coding assistant. Claude Code, Codex, GitHub Copilot and Cursor all work, and each has a free way to start — the table in step 2 compares them and links every sign-up page. This is the easiest route by a wide margin, and the one to start with: you need no API key, no endpoint and no settings. The assistant is the model. When you pick a lesson's skill, it reads the lesson, does the work and answers the check itself.

1. Get one. Any of the four. If you already use one, use that.
2. Sign in once, the way the assistant tells you to.
3. Open it in the cyber-commons folder you cloned in step 3.

That is the whole route. The run block at the foot of this page does your first lesson. If you chose Claude Code you can check it from the terminal:

claude --version      # if this prints a version, you are done
claude                # run once to sign in, if you have not

This is also how these skills are meant to be used: the agentskills.io format exists so an agent can load a SKILL.md and carry out the procedure itself.

Routes B and C are the complicated way. They are for the case where you cannot or would rather not use a commercial assistant: no account, no network, or nothing allowed to leave your machine. They are folded away just below. Open them only if you need one.

curl -fsSL https://ollama.com/install.sh | sh
ollama serve &                 # not automatic on every platform
ollama pull <a model from ollama.com/library>

export NVIDIA_API_KEY=nvapi-...          # your key, from step 2

curl -sS https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"nvidia/nemotron-3-super-120b-a12b",
       "messages":[{"role":"user","content":"Reply with the single word: ready"}],
       "max_tokens":16}'

export OPENAI_BASE_URL=https://integrate.api.nvidia.com/v1
export OPENAI_API_KEY=$NVIDIA_API_KEY
export MODEL=nvidia/nemotron-3-super-120b-a12b

# Local, free, offline after the pull, nothing leaves the machine:
export OPENAI_BASE_URL=http://127.0.0.1:11434/v1
export OPENAI_API_KEY=ollama
export MODEL=<the name you pulled>

# Or a hosted free tier — Google AI Studio gives a key with no card:
export OPENAI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai
export OPENAI_API_KEY=<your AI Studio key>
export MODEL=gemini-2.5-flash

5 · How to read a lesson page

Every page is built from the same set of sections, in the same order, and the order is part of the argument.

A page only shows the sections it has something to say in. So the list below is what can appear, not a checklist every page fills in. This page is an example: it has no Risk, no Day table and no CyberTravels scene. It is about your own computer rather than about a system somebody has to secure.

1. Risk and Control — one sentence each, at the top: what goes wrong here, and what closes it. On the lessons that are about a system.
2. The hook — a scene, before anything else. Deliberately not a summary: it is the consequence of not knowing the lesson. The summary is section 3.
3. What this lesson is — the plain description, and the Day table.
4. The framework — the concept and its diagram, plus in Functions E and F an anchor line saying which part of that function's single argument the lesson moves. This always comes before any code: teaching the how before the why is the most common way a good lesson lands badly.
5. In CyberTravels — the idea in the running case study.
6. The skill — the SKILL.md as it exists in skills/, pasted in rather than copied out, so a fix is one edit to one file.
7. Run it — the run block, the same on every page.
8. What you just proved — on the lessons that ran something, and only those. A reading lesson proves nothing and says nothing here.
9. Your turn — one input to change so a number moves. The lesson is in the difference between the two numbers, not in either one.

Deciding whether to read a lesson takes section 3 alone. If you already know the idea, section 7 alone.

A section that is present on every page regardless of whether it has anything to say stops being a heading and becomes a form. The set a given lesson renders is declared in scripts/exercises/layout.py, with the reason for each omission, and a check refuses both an empty section and prose left behind for a section that no longer renders.

The hook is always CyberTravels. It sells corporate travel, and its product is an agentic platform of four agents — a workflow agent that books and refunds, a retrieval advisor, a coding agent, and a file-system agent reading vendor documents. Alex is the product engineer who shipped it. One system, every lesson. That is a deliberate cost — a lesson could always find a sharper example of its own idea — and the payoff is cumulative: the refund limit an attacker walks past in Function B is the one a detection watches in Function E and a report counts in Function F.

The three questions every page answers

Day 0, Day 1 and Day 2 appear in a table on every lesson about the system. They are not a maturity model and not a timeline — Day 0 is not "first week". They are the three questions a practitioner asks before reading anything:

Day 2 is easy to fake and the only one a sceptical reader believes. Where a lesson produces a real number, Day 2 names it; where it does not, Day 2 says what you count instead and does not pretend. Use them like this: Day 0 decides whether to read the lesson, Day 1 is what you do, Day 2 is what you put in the update to whoever asked.

Getting this funded is the part most material leaves out, and it is hardest in organisations that cannot buy the capability. A lesson that stops at the technique gives you nothing to take to the person holding the budget.

6 · Start where your work already is

Everybody builds the system first — A0.1, then A1 and A2. After that, find your row. Open the lesson in its "start at" column and read only its Day table: those three lines are enough to decide whether it is worth an afternoon.

7 · Open source first, and then buy something

Every tool named in this commons is one you can install today without a purchase order. That is a teaching decision before it is a budget one.

You cannot specify a product you have not built a bad version of. A team that has stood up Wazuh and OpenSearch, written twenty Sigma rules and watched them fire on their own traffic can ask a vendor the two questions that matter: what does this do that my rules do not, and what does it cost to keep it true. A team that has not can only compare feature lists, and a feature list is written by the seller.

The second reason is that the gaps are the finding. Building the open-source version tells you exactly where it stops. E1.0 names two out loud: there is no open-source data-loss prevention with the maturity of the other sensors, and no product in the list at all can tell you which prompt caused a file write. Neither is visible from a datasheet, and both are the reason to buy — or to build.

8 · The four vocabularies, and which question each answers

Every lesson page carries framework labels. Four frameworks are in use across this field and they answer different questions; a finding filed under the wrong one reaches nobody.

• OWASP's two Top 10s — what can go wrong. The LLM list is application-level; the Agentic list is for systems that plan and call tools. Reach for these reviewing a feature.
• MITRE ATLAS — what an attacker did. ATT&CK's grammar applied to AI. The right lens in an incident write-up.
• NIST AI RMF — how you organise to find out. GOVERN, MAP, MEASURE, MANAGE. A NIST function in a vulnerability report is a category error.
• The EU AI Act — what you must be able to show, by article. Article 15 is accuracy, robustness and cybersecurity; Article 14 is human oversight. The only one of the four that can fine you.

A control usually needs one of each: a threat it addresses, a tactic it frustrates, a function it belongs to, and an obligation it discharges. The mapping lives in curriculum/frameworks.json and every label on every lesson page is generated from it — these are indicative mappings, not a certification.

9 · What sits behind a lesson skill

Behind each lesson skill is an audit skill — the written procedure the lesson runs, in skills/. There are 14 areas, 140 skills, 140 of them with a script, and exactly one copy of each. Eighteen audits are shared by two or three lessons, which is why a lesson skill calls one rather than carrying its own copy.

You can install a whole function at once rather than a lesson at a time, and ask what is currently linked. On a Mac or Linux:

python3 scripts/install_skills.py --all --lessons A   # this function's lesson skills
python3 scripts/install_skills.py --list              # what is linked where

On Windows, the same two with python and backslashes:

python scripts\install_skills.py --all --lessons A
python scripts\install_skills.py --list

Those are links into your clone, not copies — symlinks, or junctions on Windows — so git pull updates every tool at once and an edit you make here is live in all of them. A copy would be a fork with a friendly name: you would fix a skill once and the other copies would keep the bug while still loading and still answering.

What it changes on your computer, and how to undo it

This is worth stating plainly rather than leaving you to find out. install_skills.py does exactly one kind of thing: it creates links in your agent's own skills folder.

To remove all of it, on Windows:

python scripts\install_skills.py --all --uninstall
python scripts\install_skills.py --list

To remove all of it, on a Mac or Linux:

python3 scripts/install_skills.py --all --uninstall
python3 scripts/install_skills.py --list

The first line deletes every link. The second prints what is left, so you can see for yourself that nothing is — every line should say not present.

And to remove the commons itself: delete the cyber-commons folder you cloned, the same way you would delete any other folder. There is nothing else to undo. No installer ran. Nothing was written to your system settings or the Windows registry. Running the doctor again will show you which link folders still exist, if you want to check before or after.