controlled-generation

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.70). The skill ingests the user-supplied “source image” (URL/base64) at runtime, runs a controlNetPreprocess task, and then feeds the resulting guideImageURL (derived from that outsider-provided image) into the subsequent imageInference LLM context as controlNet[].guideImage—so if the source image text/visual content comes from an outsider, it becomes model-conditioning input.