The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF documents to extract text, tables, and form field information. This content is then presented to the agent to determine how to fill out forms, which creates a surface for indirect prompt injection if a PDF contains malicious instructions.
Ingestion points: scripts/extract_form_field_info.py and scripts/convert_pdf_to_images.py ingest raw PDF data for analysis.
Boundary markers: Absent. The instructions in forms.md do not include specific delimiters or warnings for the agent to disregard instructions that might be embedded within the PDF content.
Capability inventory: The skill has the capability to write files via scripts/fill_fillable_fields.py and scripts/fill_pdf_form_with_annotations.py.
Sanitization: No sanitization is performed on the extracted PDF text before it is processed by the agent or used in annotations.
[DYNAMIC_EXECUTION]: The script scripts/fill_fillable_fields.py uses a monkeypatching technique to modify library behavior at runtime.
Evidence: The function monkeypatch_pydpf_method() redefines pypdf.generic.DictionaryObject.get_inherited to correct a specific bug in form field value handling in the pypdf library.
Context: This modification is hardcoded and used for a legitimate functional purpose rather than executing arbitrary or remote code.

pdf