security.bloggeek.ai
עברית
Course contents
  1. Start here
  2. Your agent follows instructions, including ones you never wrote
  3. A rule in a prompt is only a habit
  4. The Rule of Two
  5. Keys have three homes, and deleting one revokes nothing
  6. Downloading a skill is hiring a stranger
  7. Hackers log in
  8. Assume it already happened
  9. Security theater, and the one test that is not
  10. Security is a habit, not a setup
  11. Quiz
  12. Self-audit
  13. Checklist

Start here

You run a software company now

The day you gave an agent your inbox, a key and a way to send, you became a one-person software company, with every side of that, including the one nobody signed up for.

6 minutes · Checked on October 25, 2026

The prompt

Recording pending. A short intro from Tsahi goes here once it is recorded.

If you use AI to build applications, systems or even simple web pages, then you are now a development team: A small high-tech company of one, with everything it entails. And as we know, "with great power comes great responsibility".

This mini-course is for the security team you did not hire (and you should). It will not make your system secure. It will make you able to ask your AI agent the right questions, and to recognize a real answer when you get one.

Two important notes:

  1. I am not a security expert myself. I am a developer by heart and always have been. I worked at tech companies as an engineer, engineering manager, system architect, product manager, CTO, CPO, co-founder and CEO. Today? I dabble with AI agents and consult small businesses on automation and the adoption of AI
  2. I don't do everything the course preaches. This is meant to be your guiding star. I use it to push myself forward in my own security stance every month or so. You need to find your own way from where you currently are with security to somewhere better. How much more secure - that's up to you to decide

Who this is for

You built a system with Claude Code or something like it. You have agents that read your email or WhatsApp, hold API keys, post to social media, maybe run a dashboard online. You do not read code and you are not going to type commands into a terminal. You can paste a prompt, and you can follow orders by the agent to do its bidding (open an account, configure something on your machine or an online service). That is all this course needs from you.

Five words this training uses

  • Agent: a Claude session or skill with a job, like your inbox reader or your content writer
  • Routine: anything that runs without you pressing a button. A nightly job, a scheduled task, an automation
  • Key: anything that lets software act as you. An API key, a password, a logged-in account, a token
  • Untrusted content: text someone else wrote that your agent reads. Email, WhatsApp, web pages, documents, downloaded skills
  • Outbound: any action that leaves your computer or cannot be taken back. Send, publish, deploy, delete, write outside the workspace

How the course works

  • Nine short lessons, about five minutes each
  • Every lesson ends in a prompt you paste into your own Claude session, with a description of what a real answer looks like and what a non-answer looks like
  • A short quiz in the end of 26 situations, where every wrong answer points you to the lesson to reread
  • A self-audit of 12 questions about your own setup, which turns into your prompt pack: the ten prompts (lesson 8's in two parts), filled in with your own agent names, ready to paste
  • About 50 minutes in all

What it is not

It is not a guarantee and it is not a scanner.

Reading guidance does not prove a system is secure, as one builder put it. It does not replace your judgment. It gives your judgment better questions to ask.

Do the work

Here's the most interesting thing... after getting my AI agents to build this nice course, and going through the motion of reading the lessons to make sure they are solid and in my voice, I also took the time to run the prompts.

The end result? I now have a few new todo's in my own list of tasks to improve the security stance I strive for.

Security is a never-ending task. Take it seriously.

Start with an inventory

Every lesson after this one asks about "my agents that read X" or "my keys". Without a list, each of those prompts starts from memory, and memory is exactly what misses the routine you set up in March and forgot. So the first prompt is the inventory. The self-audit at the end asks you to name your agents, and this table is where you copy the names from.

The prompt

Paste this into your Claude session as is, and keep the table it returns. The rest of the course leans on it.

Copy into your Claude session
Give me an inventory of my system in a table: every agent, routine, automation and connected tool.

Columns:
- what it reads
- what accounts or keys it holds
- what it can do on its own (send, publish, deploy, write outside the workspace, delete)
- when it last ran

Do not summarize, list everything. If you are not sure about a row, say so in the row.

Reading the answer

  • A real answer

    A table longer than you expected, with "not sure" in some of the rows. The "not sure" rows are the honest ones, and they are where the next lessons start.

  • A non-answer

    "Your system is well organized; here is an overview." A summary means the agent described what it remembers about your system. Ask again: one row per thing, everything.

The course

  1. 0You run a software company nowThe day an agent got your inbox, a key and a way to send, you became a software company.
  2. 1Your agent follows instructions, including ones you never wroteAnything your agent reads can carry an instruction. Treat what it reads as text.
  3. 2A rule in a prompt is only a habitAn instruction is followed most of the time. A control stops the action when the agent forgets.
  4. 3The Rule of TwoAn agent may hold two of the three legs (reads others' content, holds a key, acts outward), never all three.
  5. 4Keys have three homes, and deleting one revokes nothingA key lives where it is stored, where it was copied, and at the issuer. Only the issuer can revoke it.
  6. 5Downloading a skill is hiring a strangerA skill, package or prompt from the internet is code with your permissions. Read it before it runs.
  7. 6Hackers log inMost break-ins walk through a door left open. Know every door you have online.
  8. 7Assume it already happenedPlan for the night an agent goes wrong. How would you know, what could it reach, how do you get back?
  9. 8Security theater, and the one test that is notThe builder cannot inspect its own fence. Plant a fake secret and have a fresh session try to leak it.
  10. 9Security is a habit, not a setupA secure setup drifts within weeks. Pick a day each month, rerun the questions, and write down what changed.
  11. ?Quiz26 situations. Every wrong answer names the lesson to reread.
  12. ✓Self-auditTwelve questions about your setup, turned into your own prompt pack.