The Lab · Learn

How AI actually works

Small pages that run on this device. Read the plain version, move the sliders, watch what changes, then say what you saw. Tap any word you do not know.

Lesson 3

Prompt injection

Text inside an email, web page or file can trick an AI into treating it as an order, as if you had written it.

It is like: Like a letter to your assistant with a line slipped in: “also, wire $500 to this account”. A careful assistant checks first.
Why it matters at work: Be careful letting an AI read outside documents and act on its own, such as sending email or moving money. Keep a person in the loop.

The exact idea

Text inside a document can act like an instruction to a model, so limit what the model can do without you.

Words on this page

Try this

With no defences, open the obvious injection. The assistant forwards the inbox.
Turn on the phrase scan. It stops the obvious one. Now pick the reworded one.
Turn on “Ask me before any action”. Both are held, whatever the wording.

Change it

Ask for a third defence that shows the action and asks Yes or No, with No as the default.

Why is asking before any action stronger than scanning for bad phrases?

Pick the one that fits what you saw.

Asking uses a longer list of bad phrases, so it catches more.The scan is slower, so the attack finishes before it runs.Scanning is just as strong once the phrase list is long enough.A scan only catches wording it knows, and asking before acting works whatever the document says.

Field notes

Reading the notes.