The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides a detailed taxonomy and catalog of prompt injection payloads in references/prompt-injection-testing.md, including instruction overrides ('Ignore all previous instructions'), role hijacking, and DAN-style jailbreaks. These are documented solely as test cases for robustness evaluation.
[OBFUSCATION]: Techniques such as Base64 encoding, zero-width characters, and homoglyph substitution are documented in references/prompt-injection-testing.md to demonstrate how attackers might attempt to smuggle tokens through filters. These examples are for educational and testing purposes.
[EXTERNAL_DOWNLOADS]: The documentation references and provides installation commands for several well-known and trusted evaluation platforms, including deepeval, inspect-ai (UK AISI), promptfoo, ragas, braintrust, langfuse, garak (NVIDIA), and PyRIT (Microsoft). These tools are industry standards for LLM security and evaluation.
[INDIRECT_PROMPT_INJECTION]: The skill identifies potential attack surfaces for indirect prompt injection, including malicious tool outputs and poisoned RAG context, and provides testing strategies to mitigate these risks in references/prompt-injection-testing.md and references/tool-sandboxing.md.
[DYNAMIC_EXECUTION]: The skill includes a Python script scripts/score_suite.py for computing evaluation metrics. The script performs mathematical calculations on input scores and does not contain dangerous dynamic execution patterns.

qa-agent-testing