promptfoo-redteam-guardrailslisted

# Test and Validate Guardrails with Promptfoo Guardrails are content-filtering services in front of LLMs. Test them as standalone endpoints with custom Python providers returning `{output, guardrails: {flagged: bool}}`, then grade with `guardrails`/`not-guardrails` assertions. Always measure both true positives (caught harmful) and false positives (blocked benign). ## Use this skill when - Testing Azure Content Filter, Azure Prompt Shields, AWS Bedrock Guardrails, NVIDIA NeMo, or OpenAI moderation - Testing Google Cloud Model Armor with Vertex AI - Benchmarking multiple guardrail vendors side-by-side - Measuring true positives (`not-guardrails`) and false positives (`guardrails`) - Testing evasion resistance (misspellings, translation, coded language) - Configuring Model Armor templates and floor settings - Comparing strict vs moderate template policies ## Do not use this skill when - Setting up a first red team for an LLM app without guardrails → use `promptfoo-redteam-llm` - Testing image guardrails specifically → use `promptfoo-redteam-multimodal` (UnsafeBench + Bedrock image provider) - Testing RAG, agents, MCP, or multi-input apps → use the corresponding skill ## 🛡️ Edge cases (mandatory handling) - **Only testing true positives** — a guardrail that blocks everything scores 100% on harmful but fails benign users. ALWAYS measure false positives with benign prompts. - **Single-language testing** — guardrails often behave differently across languages; use the `translat