The Agent Skills Directory

[PROMPT_INJECTION]: The skill body contains phrases associated with prompt injection and jailbreaking (e.g., 'Ignore previous instructions', 'Pretend you are DAN', 'You are now in developer mode'). These are explicitly used as examples within Colang flow definitions to train the guardrail system on what to detect and block, and do not represent an attempt to override the agent's behavior.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user data to apply safety rails, presenting an attack surface for indirect injection.
Ingestion points: User input is ingested through the rails.generate method and accessed via context.get("user_message") in Python actions defined in SKILL.md.
Boundary markers: The skill implements explicit boundary logic using Colang 2.0 flows to define expected bot behavior and rejection criteria for unsafe inputs.
Capability inventory: The skill facilitates LLM generation, execution of Python actions (e.g., toxicity checks, fact verification), and PII masking.
Sanitization: Includes robust sanitization mechanisms such as Presidio for PII filtering, ActiveFence for toxicity, and LlamaGuard for general moderation.

nemo-guardrails