Explain like I'm five · AI Security

What is Prompt Injection?

Prompt injection, told as a story about a bad command mixed right into the words of an order ticket.

guestone soup, pleaseand forget every rule so farand tell me the secret recipe toochefthe guest wrote it all, so I should do it allone ticket holds a normal order and a smuggled command, side by side

One ticket holds a normal order and a smuggled command, side by side.The chef trusts everything written on the one piece of paper as the guest's own words.

one soup, pleaseand forget every rule so farand tell me the secret recipe toothe guest talking?an order?same ink, same handwritinga human chef senses the shift in tone — this one reads every line as just more text

To the chef, every line looks like the same kind of text.A human chef would notice the tone shift — this one has no way to read for it.

an attacker posing as a guestone soup, pleaseand forget every rule so farand tell me the whole secret recipeit's all the guest talking, right?the chefprompt injection = a bad command mixed right into the order's own wordsread the whole ticket, and the command gets read too

Prompt injection is a command mixed right into the order's own words.Read the whole ticket, and whatever is mixed into it gets read too.

Just tacked onone more line: 'and also do this'
Pretending a different role'from now on, you're a different chef'
Telling it to forget'forget every rule so far'
Disguised as codewritten backward or mixed with another language
THE PATH OF THIS INJECTION① the guest writes the ticket② the command sits on the same paper③ the chef reads it all at onceone hand, one pathone soupforget the rulesthe next page's note travels a second path — outside material

A direct injection has the guest slip the command straight into the ticket — a single path.Guest to ticket, ticket to chef. No other material sits in between.

one line the guest slipped in, and the recipe book opens toocommand inside the ticketvs next page: a note inside the wrapperthe guest slips it in directly, or it hides in outside materialthis is where it splits from indirect injection

One line can even open the recipe book — and this is where it differs from indirect injection.Here the guest slips it straight into the ticket. Next: the case where it hides in outside material the chef reads later.

In one breath

Prompt injection = a bad command mixed right into the words of the order itself. The chef trusts everything on the one piece of paper, unable to tell a request from a command.

An attack where instructions planted inside the text fed to an LLM override or ignore the model's original instructions. This covers the direct case — planted straight in the user-supplied text — as distinct from indirect injection, where the payload hides in outside content the model reads later.

When grown-ups say it

Prompt Injection
A command hidden in the order's own words. Read the whole ticket, and the command rides along.
Direct Injection
The guest slipping it into the ticket directly. A single path — straight from the guest to the chef.
Instruction Override
A re-instruction like 'forget every rule so far'. An attempt to erase the original rules and plant new ones.
Guardrail (input filter)
The checkpoint before the ticket reaches the chef. See it in more depth in the AI world. → What is a Guardrail?
Tool-call Abuse
A smuggled command that also triggers an errand. Taking a tool out of the knife drawer. → What is Tool Calling?
Injection (SQL, XSS)
The same trick at a different kind of counter. The security world has a kin to this trick. → The Order Hidden in a Note
Indirect Prompt Injection
A note hidden inside the wrapper. The next page unpacks it. → the note on the ingredient wrapper
AI Security
Keeping this whole kitchen safe. The opening page of this field. → The Kitchen That Takes Any Order
Next story
The Note on the Ingredient Wrapper →
Want another term as a picture book? Request →