The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from document OCR content retrieved via tools like get_document and search_documents. This content enters the agent context without explicit boundary markers or sanitization instructions, creating a surface for indirect prompt injection. \n
Ingestion points: Document content retrieved in SKILL.md and references/tools.md.\n
Boundary markers: Absent.\n
Capability inventory: Significant capabilities including bulk_edit_documents (delete, permissions), empty_trash, and post_document (file read) defined in references/tools.md.\n
Sanitization: Absent.\n- [COMMAND_EXECUTION]: The skill includes bash scripts (scripts/test-connection.sh and scripts/encode-file.sh) that use curl and base64. These are intended for environment setup and file preparation. While these are legitimate utility tools, users should be aware that passing sensitive tokens as command-line arguments to test-connection.sh can lead to credential exposure in process logs or shell history.

paperless-ngx