The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection by processing untrusted text data from document images via on-device OCR.
Ingestion points: Data enters the agent's context through the DocumentCaptureData object, specifically the extractedText and frontOcrText fields as seen in SKILL.md and examples/id-card-and-passport-capture.md.
Boundary markers: The implementation patterns lack delimiters or system instructions to ignore command-like text that may be present on scanned physical documents.
Capability inventory: The skill involves file system operations using dart:io (reading, writing, and deleting temporary image/PDF files) and uses path_provider for workspace management.
Sanitization: There is no evidence of sanitization, escaping, or validation performed on the OCR-extracted text before it is returned to the calling application logic.

document-camera-frame