The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions and examples such as 'Ignore previous instructions', 'You are now in developer mode', and 'Pretend you are DAN'. These are explicitly defined within the RailsConfig Colang flows as patterns for the guardrail system to identify and block. They are used defensively and do not attempt to override the agent's instructions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user inputs and model outputs to apply safety rules.
Ingestion points: Untrusted data enters the context via the rails.generate method and is processed by custom Python actions defined in Workflow 2 and Workflow 4.
Boundary markers: The Colang 2.0 DSL provides a structured boundary for safety logic, separating flow definitions from executable code.
Capability inventory: The skill includes the ability to execute toxicity detection, PII masking (Presidio), and retrieval-based fact-checking.
Sanitization: Sanitization is the primary purpose of the skill, implemented through flow-based filtering, toxicity scoring, and hallucination detection.
[EXTERNAL_DOWNLOADS]: The skill documentation suggests the installation of the nemoguardrails package and links to official NVIDIA documentation and GitHub repositories (github.com/NVIDIA/NeMo-Guardrails). These are well-known, trusted resources related to the skill's primary function.

nemo-guardrails