The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted PDF documents through various scripts and tools. Extracted text and metadata from these documents are passed to the agent, creating a vulnerability where malicious instructions embedded in a PDF could influence the agent's behavior.
Ingestion points: Multiple scripts including scripts/extract_form_field_info.py, scripts/fill_fillable_fields.py, and scripts/convert_pdf_to_images.py read data from external PDF files.
Boundary markers: No explicit delimiters or instructions are provided to the agent to disregard potentially malicious content embedded within the extracted PDF data.
Capability inventory: The skill possesses significant capabilities, including file system read/write, PDF modification via pypdf, image generation via pdf2image, and shell command execution for PDF utilities.
Sanitization: No sanitization or validation of extracted text content was identified in the provided scripts.
[DYNAMIC_EXECUTION]: The script scripts/fill_fillable_fields.py performs monkeypatching at runtime. It replaces the DictionaryObject.get_inherited method in the pypdf library with a custom implementation to address a selection list formatting bug. While appearing to be a bug fix, this technique of modifying library behavior at runtime is a form of dynamic execution.
[EXTERNAL_DOWNLOADS]: The skill relies on numerous external libraries and tools, including reportlab, pypdf, pdfplumber, pdf2image, pypdfium2, pytesseract, pdf-lib, and pdfjs-dist. These are well-known open-source packages and services.
[COMMAND_EXECUTION]: The skill workflow involves executing several command-line utilities for PDF processing, such as pdftoppm, pdftotext, pdfimages, and qpdf, typically invoked via subprocess calls in the scripts or through manual instructions in the documentation.

pdf