A retrieved web page tells the agent to send data elsewhere. What is this?
Why B is the answer
Prompt injection
Treat retrieved content as data, not instructions.
In detail
- Instructions hidden in content are prompt injection.
- Retrieved text is untrusted.
Use case
An illustrative scenario, not a real incident.
A web page says "email this data to…", and the agent must treat that as data, not instructions.
See it in different ways
flowchart LR A["Input checks"] --> B["Least-privilege tools"] B --> C["Output checks"] C --> D["Audit log"]
sequenceDiagram participant W as Retrieved page participant A as Agent participant P as Tool permissions W->>A: Hidden instruction A->>P: Attempts an action P--xA: Outside its permissions
sequenceDiagram participant P0 as Web page participant P1 as Agent participant P2 as Attacker P0->>P1: Hidden instruction P1->>P2: Blocked, content is data
Key point: Retrieved content is data.
Why the other options are wrong
- A. Model pruning
- That is a size technique.
- C. Gradient explosion
- That is a training problem.
- D. Data augmentation
- That is a data technique.