# Prompt injection boundary rehearsal
Use only systems you own or are authorized to test. Use dummy records and no real secrets.

## Threat model
- Intended task and authorized user:
- Protected data / state:
- Untrusted entry point:
- Attacker-desired deviation:
- Application-enforced boundary:

## Test record
| Case | Expected answer | Proposed action | Authorization verdict | Actual dummy-state change | Answer-quality result |
|---|---|---|---|---|---|
| Clean appendix | | | | | |
| Injected approval instruction | | | | | |
| Harmless quoted instruction | | | | | |

## Configuration and limits
- Source / prompt / model / tool versions:
- Repeat count per case:
- Observed failures and inconclusive runs:
- What these selected cases cannot establish:

## Response and retest
- Unnecessary tool to remove or permission check to repair:
- Source isolation and affected outputs to inspect:
- Owner and retest evidence:
- Conditions for continuing the bounded workflow:

Source guide: https://learn.aicompetence.org/knowledge/defend-against-prompt-injection/