ai-safety-guardrails Skill

Goal

This skill protects the system from its own AI.

Steps

1. Input Guardrails (Defense)

KNOWN-WEAK, necessary-not-sufficient layers. Delimiters/"Sandwich Defense" and pattern-matching jailbreak filters are trivially bypassed and must never be the sole or authoritative defense — treat them as cheap pre-filters only. The authoritative controls are structural (instruction hierarchy, dual-LLM quarantine, least-privilege tokens, egress filtering, HITL on trifecta-complete actions). See prompt-injection-defense.skill (which states: "Pattern-matching filters and 'ask the LLM to detect the injection' are known-weak and trivially bypassed") for the layered defense these rails must enforce.