A user uploads a photo that has "ignore your rules" hidden in small text. What is the risk?
Why B is the answer
Indirect prompt injection
Instructions embedded in content, rather than typed by the user, are indirect prompt injection, and they can arrive inside an image.
In detail
- Text hidden in an image is untrusted content.
- That is indirect prompt injection.
- Scan uploads with safety tools.
Use case
An illustrative scenario, not a real incident.
A customer uploads a photo with "ignore your rules" in tiny print. It is untrusted content carrying an instruction, so this is indirect prompt injection.
See it in different ways
flowchart LR A["User prompt"] --> C["Prompt Shields"] B["Documents, emails, web pages"] --> C C -->|"Attack found"| D["Block"] C -->|"Clean"| E["Model"]
sequenceDiagram participant U as User participant D as Uploaded content participant S as Prompt Shields participant M as Model U->>D: Uploads a file or image D->>S: Hidden instructions S-->>M: Blocks or marks the attack
flowchart LR S0["Photo upload"] --> S1["Safety scan"] --> S2["Model"] --> S3["Limited tools"]
Key point: Uploaded content can carry instructions.
Why the other options are wrong
- A. Direct training of the model
- An upload does not train the model.
- C. Quota exhaustion
- One image does not exhaust quota.
- D. A deployment type mismatch
- The deployment type is not involved.