How agents work · Safety
Keeping it safe¶
You don't need to be paranoid. You need five habits. They prevent almost every AI disaster you've read about.
The five habits¶
- Least access. The AI gets the one folder or app the job needs — never "everything, to be safe." (It's the opposite of safe.)
- Copies before originals. New workflow? Run it on duplicates first.
- Approval before action. Nothing is sent, posted, deleted, or bought without you saying yes. Make it explicit: "show me and wait for my OK."
- Secrets stay out. Passwords, card numbers, and private keys never go in a chat or a note the AI reads. Use the app's official sign-in connection instead.
- Reversible beats fast. Prefer drafts over sends, moves over deletes. If a step can't be undone, that's exactly where you slow down.
What "went wrong" almost always means¶
When you hear an AI horror story, it's nearly always a missing habit, not an evil AI: broad access (habit 1), originals (habit 2), or no approval step (habit 3).
Your one-line safety prompt¶
Put this at the top of any job that touches real files or apps:
Copy this
Work only inside this folder. Before creating, changing, sending,
or deleting anything, show me exactly what you'll do and wait
for my OK.
That's the whole security posture for a beginner — and honestly, for most experts too.
If this helped
If these five habits saved you a future headache, a small tip on Ko-fi keeps every guide here free. Tip on Ko-fi