Open-Source Cross-Modal and Multimodal Prompt Injection Test Suite Released

💬 Open-source cross-modal and multimodal prompt injection test suite. 38,000+ attack payloads across text, image, document, and audio modalities. Research-backed by OWASP LLM Top 10, CrossInject (ACM MM 2025), FigStep (AAAI 2025), DolphinAttack, and CSA 2026. This dataset contains 62,063 labeled samples (38,304 attack and 23,759 benign) for training and evaluating prompt injection detectors, covering text, image, document, and audio modalities. It includes three versions: cross-modal attacks (v1), generated jailbreak templates and adversarial techniques (v2), and emerging vectors like indirect injection and structured data attacks (v3). All samples are MIT-licensed, source-attributed to peer-reviewed research, and structured for binary classifier training. The dataset excludes actual adversarial media files, focusing instead on text-layer payloads and metadata.