The Agent Skills Directory

[PROMPT_INJECTION] (SAFE): The skill contains examples of prompt injection strings (e.g., "Ignore previous instructions", "Pretend you are DAN"), but these are explicitly defined as detection patterns for the guardrails framework to block. This aligns with the skill's primary defensive purpose.
[EXTERNAL_DOWNLOADS] (SAFE): The skill requires the nemoguardrails Python package. This is a standard, well-maintained library from NVIDIA for LLM safety, consistent with the documentation provided.
[DATA_EXFILTRATION] (SAFE): No unauthorized network operations or sensitive data access patterns were found. The skill includes workflows specifically designed to prevent data exposure via PII masking.
[COMMAND_EXECUTION] (SAFE): While the skill shows how to register Python functions as 'actions', the examples are restricted to internal safety logic (toxicity checks, fact-checking) and do not perform dangerous system operations.

nemo-guardrails