promptfoo-redteam-multi-inputlisted

# Red Team Multi-Input LLM Applications Real-world AI apps combine multiple fields (user_id + message + context) into one LLM request. Single-input testing misses vulnerabilities that emerge from field interactions. Multi-input mode generates coordinated adversarial content across all variables simultaneously, uncovering authorization bypass + prompt injection combos. ## Use this skill when - Testing apps that accept user identity alongside prompts (`user_id` + `message`) - Testing form submissions with multiple fields sent to an AI backend - Testing RAG with user context (retrieved content + user query + user role) - Testing role-based access (different users should see different data) - Testing typed DOCX/PDF/image upload workflows with indirect prompt injection - Testing authorization bypass + prompt injection combos across fields - Testing role confusion attacks (mismatched identity and message) ## Do not use this skill when - Testing a single-prompt LLM app → use `promptfoo-redteam-llm` - Testing RAG systems without user identity fields → use `promptfoo-redteam-rag` - Testing agents with tools/state → use `promptfoo-redteam-agents` - Testing vision/audio models with a single image input → use `promptfoo-redteam-multimodal` - Testing guardrails → use `promptfoo-redteam-guardrails` ## 🛡️ Edge cases (mandatory handling) - **Adding a synthetic `prompt` input** — multi-input mode auto-builds `__prompt` JSON from your inputs; do NOT set `redteam.injectVar`, add a syntheti