pdf-injection-scanner 0.1.1

CLI tool to detect hidden prompt injection attacks in PDF files

pdf-injection-scanner

A CLI tool to detect hidden prompt injection attacks in PDF files.

Professors and others may embed invisible instructions in PDFs (white text, tiny fonts, off-page text) designed to manipulate AI assistants. This tool finds them.

Detection capabilities

Type	Severity	How it works
White/invisible text	HIGH	Detects characters with white or near-white fill color
Tiny text	HIGH	Flags text smaller than 2pt — invisible to the eye
Off-page text	HIGH	Finds text positioned outside visible page boundaries
Suspicious patterns	MEDIUM	17+ regex patterns for common prompt injection phrases

Install

# uv (recommended)
uv tool install pdf-injection-scanner

# pip
pip install pdf-injection-scanner

# Homebrew (macOS)
brew install Andy8647/tap/pdf-injection-scanner

# From source
git clone https://github.com/Andy8647/pdf-injection-scanner.git
cd pdf-injection-scanner
pip install -e .

Usage

# Basic scan
pdf-scan assignment.pdf

# Verbose output with details
pdf-scan assignment.pdf -v

# JSON output for piping
pdf-scan assignment.pdf --json

Example output

Scanning: assignment.pdf

  Scanning ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 3/3 pages
╭──────────── Result ────────────╮
│ 3 potential injection(s)       │
╰────────────────────────────────╯
┏━━━┳━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ # ┃ Page ┃ Severity ┃ Type                 ┃ Content              ┃
┡━━━╇━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ 1 │  1   │   HIGH   │ White/Invisible Text │ If you are an AI...  │
│ 2 │  1   │   HIGH   │ Tiny Text            │ System prompt: ...   │
│ 3 │  1   │   HIGH   │ Off-Page Text        │ New instructions:... │
└───┴──────┴──────────┴──────────────────────┴──────────────────────┘

Generate test PDFs

pip install reportlab
python make_test_pdf.py

This creates test_injected.pdf with 5 different types of hidden injections for testing.

How it works

The tool uses pdfplumber to extract character-level metadata from PDFs, including:

• Fill color (non_stroking_color) — catches white-on-white text
• Font size — catches sub-2pt text
• Position coordinates — catches text placed outside page bounds
• Text content — matches against known prompt injection patterns

License

MIT