The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses session transcript files located in ~/.claude/projects/<project-slug>/*.jsonl. These logs contain the full history of user messages and agent responses, presenting a significant data exposure risk.\n- [COMMAND_EXECUTION]: The skill executes multiple shell commands, including sips for image rotation and a Python script for parsing session transcripts. This script is generated and executed dynamically based on instructions in the markdown.\n- [DATA_EXFILTRATION]: The skill uses curl to upload local image assets to remote URLs. While intended for processing, this establishes a network path for data leaving the local environment.\n- [PROMPT_INJECTION]: The skill has an indirect prompt injection surface when processing user photos.\n
Ingestion points: User-provided images (URL or local path).\n
Boundary markers: The prompt templates use thematic sections but lack explicit instructions for the model to ignore potential instructions embedded within the image content.\n
Capability inventory: The skill utilizes generate_image_edit, upload_asset, and remove_background tools, alongside shell execution capabilities (sips, curl, python3).\n
Sanitization: No filtering or sanitization of image content or metadata is performed before the generation process.

voxel-it