The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from untrusted PDF documents, which could contain malicious instructions designed to influence the agent. 1. Ingestion points: PDF files are read using extract_form_field_info.py, convert_pdf_to_images.py, and pdftotext. 2. Boundary markers: The workflow in forms.md requires visual analysis and JSON mapping, creating a separation between content and instructions. 3. Capability inventory: The skill can write files and execute local Python scripts. 4. Sanitization: Safety relies on the agent's validation of field mapping rather than automated sanitization.
[DYNAMIC_EXECUTION]: The script fill_fillable_fields.py modifies the behavior of the pypdf library at runtime to address a specific bug in field handling. This modification is restricted to a local library call and does not involve executing code from remote or untrusted sources.

pdf