PDFender

PDFs supplied to generative AI may contain AI-directed instructions that are difficult for a person to notice. PDFender provides human-review clues for finding this form of indirect prompt injection.

The app examines PDF text data and drawing operations to compare extracted text with the page as a person sees it. Rather than using OCR to read images, it checks text color, size, opacity, placement, and paint order for visually hidden text.

Main Features

• Detects text that blends into its background, abnormally small or transparent text, and text covered by shapes or images.
• Reports invisible Unicode characters and instruction-like text in PDF metadata as review candidates.
• Provides navigation to findings, detection signals, and scores to help compare results with the original PDF.
• Loads and analyzes PDFs entirely in the browser. PDF contents, extracted text, and findings are never sent to an external service.

Findings are review clues, not proof that a PDF is safe or contains an attack. False positives and false negatives are possible. Image-only text and some other content are outside the inspection scope.

Detection Guide

The guides explain how the detection methods work and where their limits lie.

• PDFender detection methods and limitations (English)
• PDFender：検出方式と制約のガイド（日本語）