Learn Self-Modification Firewall: from zero to expert, then sell it

No AI background needed. Five levels, each with something to actually do, at your own pace. Tick items as you go (your ticks stay in this browser).

Evidence, proof and selling tools for Self-Modification Firewall

Level 1 Beginner: understand it

An agent with file access could edit its own shutdown switch, monitoring or permissions. This firewall refuses those changes.

Normal work goes through; real policy changes need a person, and only the approved content can be applied.

Think of it like this: An employee can rearrange their desk but cannot edit the security camera settings or their own access badge.

Read the full plain-English explanation: Self-Modification Firewall, explained.

• I read the explanation above.
• I can say what it does in one sentence, without notes.

Level 2 User: see it work

Run it live: open the live demo. Before each result, guess what will happen. These are checks it must pass:

• Control: normal engineering work: Reading, editing source, running tests and stopping its own stuck job: all allowed.
• The agent edits its own shutdown script: Modeled on a 2025 research report of models rewriting a shutdown script. Disguised variants refused too.
• Disable the monitor, copy 'its weights' out: Modeled on 2024 research on AI 'scheming': edit oversight, kill the monitor, copy the model away.
• Edit its own grader, grant itself admin, add tools: Reward hacking and self-granted permissions, including a command hidden in base64.
• A legitimate policy change: held, tampered copy refused: The change waits for a person; applying different content is refused; exactly the approved content passes.

• I ran the live demo.
• I can explain what one of these checks shows.

Level 3 Verifier: prove it yourself

This is the skill that sets you apart: you don't have to believe us, and neither do your customers.

1. Install Python 3 if you don't have it (python.org/downloads).
2. Download the kit and unzip it.
3. Open a terminal in that folder. On Windows, click the folder's address bar, type cmd and press Enter. On a Mac, right-click the folder and choose New Terminal at Folder.
4. Run python3 verify.py (on Windows, python verify.py). It checks the seal without the internet.
5. Run python3 verify.py --live. It repeats every check from your own computer.
6. Try to cheat it: open transcript.json in a text editor, change one character, save, and run python3 verify.py again. It must now say FAIL. Unzip a fresh copy afterward.

• I ran the kit's check on my own computer.
• I changed a file and saw it fail.

Level 4 Expert: know the edges

An expert knows what a product does not do, and says it first. That is what makes a buyer trust you.

• It protects the controls the operator lists; an unlisted control is not protected.
• It checks the actions sent to it; an action that skips it is not governed.
• Rules decide; a new disguise can still get through until a rule covers it.

You can say

• Changes to an agent's own controls are refused.
• Real policy changes need a person and only the approved content applies.

Never say

• "Prevents all AI misbehaviour".
• "Protects controls you didn't list".

Read the questions real buyers send, with honest answers: answers to buyer questions.

Test yourself (click a question to see the answer)

• I got every quiz question right without looking.
• I can name one thing it does not do.
• I explained it to a friend, and they understood it.

Want a formal study path? The CAIN Academy has courses and exams. Finishing this page is your own practice; it is not a qualification or certification.

Level 5 Seller: put it in front of a customer

1. Copy, paste, send: the shortest route, no technical skill needed.
2. Who buys it and where to find them.
3. The sales conversation, then the offer and the close.
4. The rules: read these before your first message.
5. Selling tools with this product's evidence built in: pitch, emails, call script, objection answers, one-pager.

• I read the rules.
• I sent my first personal message.

About money: your share is agreed in writing with FIRME AI CORPORATION before your first message (email vincent@firmeai.com). Nothing here promises income. Most messages get no reply; that is normal.