pdf-vision
Pass
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script 'scripts/ocr_columns.py' executes a local Swift utility ('scripts/ocr.swift') via 'subprocess.run' to process image segments. This uses a secure argument list format and is part of the intended OCR workflow.
- [EXTERNAL_DOWNLOADS]: Extracted text content is sent to external embedding services (Google Gemini, OpenAI, or Google Vertex AI) for vector generation via API calls. These are legitimate operations for a RAG skill and target well-known service providers.
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted data from PDFs which could contain malicious instructions.
- Ingestion points: Document text is ingested via 'pdfplumber' in 'scripts/pdf_check.py' and via Apple Vision OCR in 'scripts/ocr.swift'.
- Boundary markers: None identified; the skill does not wrap extracted text in delimiters or provide explicit instructions to the agent to treat the content as untrusted data.
- Capability inventory: The skill includes file writing to the user's home directory ('~/.cli-jaw/vectors/'), network API calls for embeddings, and subprocess execution for local OCR scripts.
- Sanitization: No evidence of text sanitization or validation was found for the content extracted from PDF documents.
Audit Metadata