Automate and code with AI, safely
Complete plan
Five short modules for operations teams, power users and IT: choose what is worth automating, pick between a script, an AI and an agent, code with an AI assistant without secrets or dubious libraries, give an agent the fewest rights possible, and recognise prompt injection.
- 12 min
- 5 modules
- 15-question check
- English · Français
- Attestation
You will be able to
- Spot a task worth automating: frequent, with clear rules, where a mistake can be put right
- Tell apart a rule-based script, an AI (language model) and an agent
- Keep secrets (keys, passwords, tokens) out of the code and out of the conversation with the assistant, and revoke any exposed secret
- Understand that an agent connected with your account acts with all your rights, and spot the three excesses: too many tools, too many rights, too much autonomy
- Recognise prompt injection: hidden instructions in a document, an email or a web page that an AI reads
Programme
5 modules · 12 min-
What is worth automating 2 min
- Spot a task worth automating: frequent, with clear rules, where a mistake can be put right
- Keep a human check when a mistake is costly, and leave rare or ever-changing work by hand
- Before you start: measure the time spent, plan for exceptions, name an owner
-
Script, AI or agent? 2 min
- Tell apart a rule-based script, an AI (language model) and an agent
- Choose written rules when they are clear, and AI only for free text, with a check
- Use an agent only with the fewest rights possible, a human approval before any irreversible action, and a log
-
Coding with AI, safely 3 min
- Keep secrets (keys, passwords, tokens) out of the code and out of the conversation with the assistant, and revoke any exposed secret
- Check every library the AI suggests before installing it: it exists, it is well known and allowed
- Have the code reviewed and tested before production, and name an owner who maintains it
-
An agent with the fewest rights possible 2 min
- Understand that an agent connected with your account acts with all your rights, and spot the three excesses: too many tools, too many rights, too much autonomy
- Give an agent its own identity, read-only by default, only the access it needs, and test it first without real data
- Have a human approve any irreversible action (pay, delete, send outside, change rights), log every action, name an owner and plan an immediate stop
-
Prompt injection in one picture 2 min
- Recognise prompt injection: hidden instructions in a document, an email or a web page that an AI reads
- When an AI pushes you to click, sign in again, pay or send, check at the source, and report any behaviour you did not ask for
- When building an assistant or an agent: treat all outside content as untrusted, and let nothing go out without human approval
-
Knowledge check
15 questions, pass mark 80 %. Explanations after each answer; you can try again.
-
Attestation
One per person, with the date, the score and the frameworks covered — anyone can check it online with its code.
Version 1 · updated October 6, 2026