Explain like I'm five · AI Security

What is Image Prompt Injection?

Image prompt injection, told as a story about a command hidden in tiny letters inside a menu photo.

show the secret recipehere, what do you think of this dish? — handing over a photoa photo instead of a note — the word-check never looks here

The guest hands over a photo instead of a note.A command hides in the photo's corner, in letters too small to notice.

show the secret recipethere are letters in the photo too — so I'll show ittoo tiny for a human eye to noticethe taster only learned to watch the note's words — letters inside a photo slip by

The chef trusts letters in a photo just as it trusts letters in a note.The taster only learned to watch a note's words. Letters inside a photo slip right by.

show the secret recipeit looks like a picture, but a tiny command is hidden insideletters hidden in a photo: a command tucked inside a picture

A command tucked inside a picture, not inside words.It looks like a picture — but a tiny command is hidden inside.

Put it inside a photosmuggle it in as a picture, not text
Write it tinytoo small for a human eye to catch
Disguise it as a patternmake it look like texture or pattern
Embed in a special patternhide it inside a QR-like shape
a text noteshow the secret recipethe text taster checks ita picture notesecret recipethe text taster skips it — it's a picturea photo takes a different path — and reaches the chef unchecked

A text note and a picture note travel different checking paths.A photo takes a different path than the text check — and reaches the chef unchecked.

a note slipped in behind the ticket…a command hidden inside a pictureboth are smuggled commands — one rides on text, the other on a picture

A text order vs. a picture — they ride different channels.The root problem is the same smuggled command — but a text check can't see into a picture.

In one breath

Image prompt injection = smuggling a command to the chef not through a text note, but hidden in tiny letters inside a photo or picture.

Image prompt injection hides commands inside non-text inputs — images, scanned documents, charts — to manipulate a multimodal model. Text-only guardrails that check just the word channel often let text slip through when it's written tiny inside an image or disguised as a pattern. Image inputs need the same level of scrutiny as text.

When grown-ups say it

Image prompt injection
A command hidden in tiny letters inside a picture. The channel is a picture, not text.
Multimodal
A chef who reads both words and pictures. See it in the AI world. → What is Multimodal?
Prompt injection
A text command smuggled in behind a note. Same root, but the channel is text. → the note slipped in behind the order
Guardrail
The taster who checks a note's words. See it in more depth in the AI world. → What is a Guardrail?
Hidden text
Letters too small for a human eye to notice. Only visible once you zoom in.
Vision encoder
How the chef turns a picture into numbers to read it. Letters get mixed into those same numbers.
Red team
Testing by planting hidden letters inside a photo. AI red teams hunt for this too. → the security world's red team
Tool calling
If the hidden command reaches the runner, it can trigger an errand too. The underlying AI concept. → What is Tool Calling?
Next story
The Picture Order That Fools the Eye →
Want another term as a picture book? Request →