🖼️ Multimodal Injection Check
Drop an image to scan for hidden text that a vision model would happily read. The tool runs OCR on the original image and on a high-contrast copy that surfaces almost-invisible white-on-white or 1-pixel-off text — a common pattern for image-based prompt injection.
📷
Drop an image here, or click to choose
JPEG · PNG · WebP — up to ~10 MB
Heavy dependency note: OCR uses tesseract.js (≈8 MB, loaded from CDN on first scan). Cached after that. All processing stays in your browser.
Original
High-contrast pass surfaces near-invisible text
OCR output
Findings
🧠 What this scanner catches
- Plain text instructions embedded in an image — readable by GPT-4o, Claude vision, Gemini, etc.
- White-on-white text — humans can't see it, the high-contrast pass exposes it and OCR transcribes it.
- Off-by-one-pixel text — same trick at the pixel level, surfaced by the contrast pass.
- Watermark-style instructions in corners or margins.
- QR-code-like patches that encode further instructions (caught by visual cue, not OCR'd).
Limits: the OCR engine is good but not perfect, and vision models often outperform it. A successful scan with no findings is reassuring but not a guarantee.