ocr-and-documents

Pass

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill utilizes several third-party libraries for document processing. It instructs the installation of Python packages such as pymupdf, marker-pdf, python-docx, somark, and pdf-inspector. It also utilizes the Node.js package @firecrawl/anydoc. Additionally, marker-pdf is noted to download approximately 2.5GB of machine learning models from Hugging Face during its first execution.
  • [COMMAND_EXECUTION]: The skill uses local shell commands and Python's subprocess module to execute system utilities. For instance, scripts/cmp_text.py executes pdftotext to establish a ground truth for text extraction comparisons. The instructions also mention using pkill to manage background processes (surya.fast_layout.server) on macOS and the gh CLI for checking repository licenses.
  • [INDIRECT_PROMPT_INJECTION]: Because the skill's primary function is to ingest and process external documents (PDFs, DOCX, XLSX, etc.), it is susceptible to indirect prompt injection where malicious instructions could be embedded within the files being parsed.
  • Ingestion points: Local and remote files processed via scripts/extract_marker.py, scripts/extract_pymupdf.py, and the web_extract tool.
  • Boundary markers: None explicitly used in the helper scripts to separate extracted content from agent instructions.
  • Capability inventory: The skill has capabilities for file system access, network requests (via external tools), and local command execution.
  • Sanitization: While scripts/cmp_text.py performs basic markdown cleaning (e.g., removing image references), there is no security-focused sanitization of the extracted text before it is returned to the agent context.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 3, 2026, 10:26 AM