The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection through document processing.
Ingestion points: In forms.md, the agent is instructed to convert user-supplied PDFs into PNG images and then visually analyze those images to identify form fields and their purposes.
Boundary markers: No explicit boundary markers or instructions to disregard embedded text/commands within the processed images are provided.
Sanitization: There is no evidence of sanitization or filtering of the content extracted from the PDFs before it is used to influence the agent's next steps.
Capability inventory: The skill has access to local script execution, file system operations, and image processing, making it vulnerable to malicious instructions embedded within a PDF designed to hijack the agent's workflow during the 'Visual Analysis' phase.
[COMMAND_EXECUTION]: The skill relies on the execution of several local utility scripts to perform its functions.
Instructions in forms.md direct the agent to run multiple Python scripts (e.g., extract_form_field_info.py, convert_pdf_to_images.py, fill_fillable_fields.py) via the command line.
These scripts take user-provided filenames and JSON data as arguments, which could be exploited if the scripts do not properly sanitize input or if they are manipulated to execute arbitrary code.

pdf