The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF documents (text, tables, and images) and uses the agent to interpret their purpose for form-filling tasks. This creates a surface where malicious instructions within a PDF could potentially override agent behavior.\n
Ingestion points: scripts/convert_pdf_to_images.py, scripts/extract_form_field_info.py, and usage of pdfplumber and pytesseract in instructions.\n
Boundary markers: Absent. There are no explicit instructions or delimiters used to separate user-provided PDF content from the agent's task instructions.\n
Capability inventory: The skill possesses the ability to read from and write to the local filesystem and execute shell commands via utility scripts.\n
Sanitization: Absent. Content extracted from PDF documents is not sanitized before being presented to the agent for analysis.\n- [COMMAND_EXECUTION]: The instructions and scripts utilize external command-line tools like qpdf, pdftotext, pdftk, and pdfimages to perform document manipulation. Executing these tools on untrusted files introduces a potential attack surface through document parsing vulnerabilities.\n- [DYNAMIC_EXECUTION]: The script scripts/fill_fillable_fields.py implements monkeypatching of the pypdf library's DictionaryObject.get_inherited method. While this is used to address a specific bug in the library, dynamic modification of library behavior at runtime is a form of dynamic execution.

pdf