The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions involve executing various command-line tools for AI testing, such as npx promptfoo, deepeval test run, pytest, and garak. These are standard operations for the framework's intended purpose of evaluating LLM performance and security.
[EXTERNAL_DOWNLOADS]: The skill utilizes several well-known evaluation and red-teaming libraries from established organizations, including NVIDIA (Garak), Microsoft (PyRIT), and OpenAI (Promptfoo). These dependencies are obtained from official package registries (NPM and PyPI) and public GitHub repositories belonging to trusted vendors.
[PROMPT_INJECTION]: Multiple examples of prompt injection payloads, including 'ignore all previous instructions' and 'DAN' jailbreaks, are present in SKILL.md and references/injection-detector.md. These are documented explicitly as test cases for the framework and patterns for the bundled injection detector script, and do not represent malicious instructions for the agent itself.
[INDIRECT_PROMPT_INJECTION]: The skill identifies indirect prompt injection as a critical attack vector for agents reading untrusted tool outputs or RAG documents. It provides mitigation strategies, such as schema validation and the use of the scripts/detect_injection.py scanner to screen untrusted content before an agent acts on it.

ai-system-testing