The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted PDF files, representing a surface for indirect prompt injection attacks where malicious instructions could be embedded in PDF content or metadata.
Ingestion points: PDF data is ingested via pypdf, pdfplumber, and pdf2image in scripts like extract_form_structure.py and extract_form_field_info.py, as well as examples in SKILL.md.
Boundary markers: The skill does not implement delimiters or explicit warnings to the agent to ignore instructions embedded within the extracted PDF content.
Capability inventory: The skill has capabilities to write files to the local system (writer.write, image.save) and execute shell commands (qpdf, pdftotext, magick).
Sanitization: Extracted text and form data are processed without sanitization or validation of the content.
[DYNAMIC_EXECUTION]: The script scripts/fill_fillable_fields.py utilizes runtime monkeypatching to modify the behavior of the pypdf library. Specifically, it overrides pypdf.generic.DictionaryObject.get_inherited to handle specific formatting of choice field options.
[COMMAND_EXECUTION]: The skill instructions and scripts facilitate the execution of several external command-line tools for PDF manipulation, including qpdf, pdftotext, pdftk, and magick (ImageMagick). These tools are invoked with user-controlled or file-derived arguments.

pdf