Guardrails & Safety Skill

Purpose

Provides a comprehensive, production-grade safety framework for AI agent deployments. This skill addresses the full spectrum of guardrail layers — from ingress input validation through internal logic authorization to egress output sanitization. It defines detection algorithms for prompt injection attacks, hallucination scoring pipelines, PII redaction engines, tool-call permission matrices, retrieval pipeline security hardening, programmable policy enforcement via NeMo Guardrails and Guardrails AI, and multi-layer content filtering stacks. The framework is designed to be composable: each guardrail layer operates independently but can be chained into a unified enforcement pipeline with centralized monitoring and alerting.

Core Principles

1. Defense in Depth: Never rely on a single guardrail layer. Stack input, logic, output, and retrieval guardrails to create overlapping defense zones.
2. Fail-Closed by Default: When a guardrail check is inconclusive or produces an error, the system must reject the request rather than allow it through.
3. Least Privilege Execution: Agent tool invocations must be scoped to the minimum permissions required. Every tool call passes through an authorization matrix before execution.
4. Transparent Auditability: Every guardrail trigger, bypass, and override must be logged with full context for post-incident forensics and compliance reporting.
5. Adversarial Resilience: Design guardrails to withstand known attack taxonomies including direct injection, indirect injection, jailbreaking, and prompt leaking.

Agent Protocol