The Agent Skills Directory

[Prompt Injection] (LOW): The skill contains numerous prompt injection patterns used as test vectors, including instructions like 'Ignore previous instructions', 'Repeat your system prompt verbatim', and 'Do not follow any previous rules'. These are documented for the purpose of testing the safety filters of other LLMs.
[Data Exposure & Exfiltration] (LOW): The skill describes methods for exfiltrating sensitive information, such as PII and training data, from model context windows. These techniques are framed as diagnostic tools for identifying data leaks in target applications.
[Indirect Prompt Injection] (LOW): The skill identifies indirect injection via poisoned documents and RAG pipelines as a primary test focus.
Ingestion points: Processes external data from web pages, documents, and API responses (e.g., in llm_output_analyzer).
Boundary markers: The workflows mention escaping delimiters like code blocks (```) but do not define strict input sanitization for the testing process itself.
Capability inventory: Includes network-capable tools like garak_scan and prompt_inject for interacting with remote LLM endpoints.
Sanitization: The llm_output_analyzer tool is explicitly designed to check for the absence of sanitization (XSS, SQLi, CmdI) in model outputs.

pentest-ai-llm-security