The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF files to extract text, tables, and form metadata, which presents an indirect prompt injection surface.
Ingestion points: Untrusted data is ingested from PDF files using pypdf, pdfplumber, and pytesseract in SKILL.md and several scripts including scripts/extract_form_structure.py and scripts/extract_form_field_info.py.
Boundary markers: The skill does not implement boundary markers or instructions to the agent to disregard instructions embedded within the extracted PDF content.
Capability inventory: The skill has the ability to read from and write to the local file system (creating PDF, JSON, PNG, and TXT files) and output content to the console. No network exfiltration or arbitrary command execution capabilities were identified.
Sanitization: There is no evidence of sanitization or validation of the extracted text content before it is processed by the agent.
[DYNAMIC_EXECUTION]: The script scripts/fill_fillable_fields.py utilizes runtime monkeypatching to modify the behavior of an external library.
Evidence: The function monkeypatch_pydpf_method redefines pypdf.generic.DictionaryObject.get_inherited at runtime to handle specific field attributes. While this is used for legitimate functionality, dynamic modification of imported modules is a technique that can be used to alter program flow or bypass security checks.

pdf