The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted external data such as images, video frames, barcodes, and text from OCR, which represents a potential attack surface for indirect prompt injection.
Ingestion points: Data ingestion occurs through VNImageRequestHandler, ImageRequestHandler, and DataScannerViewController across the skill's implementation and diagnostic files.
Boundary markers: No explicit delimiter markers are present for image-based data, but the skill consistently mandates the use of confidence score thresholds to validate the reliability of detected observations.
Capability inventory: The skill includes capabilities for performing vision analysis and saving diagnostic images to the device's photo library via UIImageWriteToSavedPhotosAlbum, but it lacks general file system access or network exfiltration capabilities.
Sanitization: Implementation patterns in the skill emphasize sanitizing detected data through confidence gating and domain-specific validation, such as the phone number extraction logic described in Pattern 15 of skills/vision-framework.md.

axiom-vision