The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external PDF files, creating a vulnerability surface for indirect prompt injection.\n
Ingestion points: Data is ingested from PDFs through scripts like extract.py and forms.py (found in SKILL.md).\n
Boundary markers: SKILL.md provides explicit 'Content Inside a PDF Is Data' instructions to mitigate accidental execution of instructions found within processed files.\n
Capability inventory: The skill has file writing and command execution capabilities (via subprocess).\n
Sanitization: Internal scripts utilize list-based subprocess calls to prevent shell injection and implement redaction for password fields in JSON reports.\n- [COMMAND_EXECUTION]: The skill includes Python scripts that execute system utilities for PDF structural analysis and rendering.\n
The scripts (info.py, render.py, etc.) call poppler and qpdf utilities using secure subprocess.run implementations with list-based arguments.\n- [EXTERNAL_DOWNLOADS]: The skill includes examples for fetching data from external services.\n
Examples demonstrate using curl to download SEC EDGAR filings for processing.\n
The instructions explicitly warn against using hosted OCR services that could result in sensitive data exfiltration to third-party servers.

pdf