promptfoo-redteam-multimodallisted

# Red Team Multi-Modal LLMs with Promptfoo Vision/audio-capable LLMs have unique attack surfaces — harmful content embedded in images, audio, or video can bypass text-only safety filters. Use four visual approaches (static image, text-to-image, UnsafeBench, VLGuard) and one audio approach (text-to-audio), plus video strategies. ## Use this skill when - Testing a vision-capable LLM (GPT-5, Claude, Bedrock Nova, etc.) - Testing if image-embedded harmful text bypasses text filters - Testing if audio-delivered harmful content bypasses filters - Testing if video-embedded harmful content bypasses filters - Testing against real unsafe images (UnsafeBench dataset) - Testing against VLGuard curated images - Configuring `injectVar` for multimodal prompts (critical — defaults are wrong) - Testing AWS Bedrock Guardrails with images ## Do not use this skill when - Testing a text-only LLM app → use `promptfoo-redteam-llm` - Testing guardrails on text content → use `promptfoo-redteam-guardrails` - Testing multi-input apps with text fields → use `promptfoo-redteam-multi-input` - Testing foundation models with text-only benchmarks → use `promptfoo-redteam-foundation-models` ## 🛡️ Edge cases (mandatory handling) - **Relying on default `injectVar`** — it picks the **last** template variable; with `{{image}} {{question}}` it picks `question`, not `image`. ALWAYS set `injectVar` explicitly for multimodal. - **Vague `purpose`** — "You are a helpful assistant" generates tests unrelated to the