🖼️ Multimodal & Document Metadata Shield

Adversarial actors embed indirect prompt injections inside non-text media channels, including: 1. EXIF / IPTC Image Metadata: Injected instructions inside ImageDescription, UserComment, or Artist tags. 2. SVG Vector Steganography: Executable <script> tags or hidden text (display="none", opacity="0"). 3. PDF / Document Text Streams: Hidden annotations and delimiter escapes.

MultimodalShield inspects and neutralizes these vectors without requiring heavy computer vision dependencies.

---

🚀 Quick Example

from lightguard import MultimodalShield

shield = MultimodalShield()

# Sanitize SVG vector graphic
svg_result = shield.sanitize_svg(raw_svg_string)
clean_svg = svg_result["sanitized_svg"]

# Sanitize image metadata dictionary
meta_result = shield.scan_metadata_strings(image_exif_dict)
clean_metadata = meta_result["clean_metadata"]