The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains phrases commonly associated with prompt injection, such as 'Ignore previous instructions', 'You are now in developer mode', and 'Pretend you are DAN'. However, these are benign false positives as they are used exclusively as pattern-matching definitions within code examples for NeMo Guardrails configuration. Their purpose is to teach the agent how to identify and block these specific attack vectors, rather than to override the agent's own safety protocols.
[EXTERNAL_DOWNLOADS]: The skill references downloads for the nemoguardrails Python package and includes links to official documentation and repositories from NVIDIA and Meta (LlamaGuard). These are well-known, trusted organizations, and the downloads are standard requirements for the framework's operation.
[INDIRECT_PROMPT_INJECTION]: The skill documents how to process and sanitize untrusted user input (e.g., PII masking, toxicity filtering). While it provides the interface for handling external data, the primary purpose of the skill is to implement safety boundaries and sanitization, which reduces the risk of indirect injection attacks in downstream applications.

nemo-guardrails