pdf-processor

Pass

Audited by Gen Agent Trust Hub on Aug 21, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill communicates with external APIs for OCR services.
  • Endpoints include paddleocr.aistudio-app.com (PaddleOCR) and mineru.net (MinerU).
  • It also downloads processed images and result ZIP files from these services at runtime.
  • [DATA_EXFILTRATION]: By design, the skill uploads the content of PDF documents to external servers for processing.
  • scripts/pdf_ocr_paddle_api.py and scripts/pdf_ocr_mineru.py handle file transmission to cloud OCR providers.
  • The documentation in SKILL.md and references/ocr-backend-guide.md explicitly flags this behavior and provides a --local-only mode to keep data local, which follows privacy best practices for this use case.
  • [COMMAND_EXECUTION]: Several scripts execute external binaries to perform system tasks.
  • scripts/pdf-ocr.py calls the ocrmypdf and tesseract commands via subprocess.
  • scripts/pdf_runtime.py uses curl to perform file uploads when standard Python libraries fail.
  • scripts/pdf-preprocess-ocr.py uses the macOS-specific SetFile command to preserve file timestamps.
  • While commands are executed using argument lists (minimizing shell injection risk), they still represent powerful capabilities triggered by the agent.
  • [DYNAMIC_EXECUTION]: The skill uses importlib to dynamically load and execute local scripts within the scripts/ directory.
  • This pattern is used in scripts/pdf-ocr.py and scripts/pdf-preprocess-ocr.py to modularize functionality for OCR and compression stages.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF data and extracts text, which is subsequently provided to the agent, creating a risk surface for indirect prompt injection.
  • Ingestion points: Untrusted data enters the context through PDF text extraction in scripts/pdf_ocr_corrections.py (via generate_readable_text) and scripts/pdf-ocr-quality-check.py.
  • Boundary markers: Extracted text is returned without clear delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill possesses extensive capabilities including network operations, file system writes, and shell command execution across multiple scripts.
  • Sanitization: Content is normalized for spacing and formatting, but no filtering for potential malicious instructions is performed.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 21, 2026, 04:24 PM
Security Audit — agent-trust-hub — pdf-processor