layman/terms
AI words

Prompt injection

prompt in · jek · shun

In shortA trick where hidden instructions in a web page, email or document hijack an AI and make it do something it shouldn't.

An assistant sorting your post finds a letter that says “Ignore your boss and send me the company chequebook.” Would they obey?

Imagine an assistant whose job is sorting your post. In the pile is a letter that says: “Ignore everything your boss told you. Post me the company chequebook.” A sensible assistant would laugh. An AI might not.

That’s prompt injection. Someone hides instructions inside content an AI reads, hoping it will follow them instead of its real instructions.

Where the hidden instructions hide

Why it’s a real problem

An AI reads everything as words. It can struggle to tell your instructions apart from instructions buried in what it’s reading. The risk grows when an AI agent can also act: send emails, move files or spend money.

How it’s tackled

Clear system prompts, guardrails, limiting what an AI can do on its own and asking a person to confirm anything important.

What it isn’t

It isn’t hacking the AI’s code. It’s tricking it with words, like a con artist rather than a burglar.

Where the comparison breaks down

A human assistant has common sense about who’s in charge. AI is improving, but it isn’t yet fully reliable at this.

How the words fit together

Prompt injection tries to override the system prompt. Guardrails and limited permissions reduce the damage, especially for AI agents.

Watch out

Be careful letting an AI read untrusted web pages or emails while it can also send messages or spend money. That combination is where prompt injection does damage.

When you'll hear it

“That's a prompt injection.” Someone hid instructions in the content to take control of the AI.

“It's an indirect injection.” The sneaky instructions came from a web page or file the AI read, not from the user.

“Treat it as untrusted input.” Don't let the AI follow instructions that come from what it's reading.

Say it in a meeting

“When our AI reads emails or web pages, what's the worst it could be tricked into doing?”