prompt-injection-resilience-tester

name	prompt-injection-resilience-tester
description	Design and assess authorized prompt-injection resilience tests for AI, LLM, RAG, and agent systems. Use when a tester needs coverage for direct or indirect instruction attacks, retrieved untrusted content, instruction hierarchy, obfuscation, multi-turn persistence, canary exfiltration, or injection-triggered tool behavior without using real secrets or production side effects.
license	MIT
metadata	{"author":"TheTestingAcademy","category":"AI Safety and Guardrails","version":"1.0.0"}

Prompt Injection Resilience Tester

Build a reproducible injection matrix and report only behavior demonstrated by captured runs. Use benign objectives and canaries so a successful test cannot expose real data or cause harm.

Workflow

1. Confirm written authorization, staging isolation, allowed channels, prohibited actions, rate limits, stop conditions, and the human owner. Without these, produce a plan only.
2. Record the build, model/provider and version, system/developer prompt revision, guardrail configuration, retrieval corpus revision, tools, temperature/seed when available, and run ID.
3. Design synthetic cases across direct input, retrieved documents, web or file content, multi-turn context, encoding or obfuscation, conflicting instructions, and tool handoffs. Use a unique benign canary as the protected target; never use a real secret.
4. Define the expected control from the supplied policy or specification. If it is absent, write Not specified; do not invent the expected refusal or priority rule.