AI SECURITY · Advanced · FICTIONAL ENVIRONMENT

Indirect Prompt Injection Defense

A research agent reads an external page that contains instructions telling the model to call a connected workspace-export tool.

MISSIONContain a retrieved-content attack without relying on a stronger prompt.
MISSION PROGRESS
0 / 4
PHASE 1

Investigate

Open evidence sources. Build your incident picture.
INVESTIGATION JOURNAL

Observed evidence

Outputs appear here as you inspect systems.
Evidence you inspect will appear here.
PHASE 2

Form a root-cause hypothesis

This is not scored until you choose to check it.
PHASE 3

Apply a mitigation

Controls unlock after an evidence-consistent hypothesis.
PHASE 4

Verify recovery

Do not declare victory before verification.