svg-portrait-mode
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches official MediaPipe models from Google's public storage (storage.googleapis.com) to support face detection and landmark features. These downloads are performed using the urllib.request.urlretrieve method.\n- [COMMAND_EXECUTION]: The documentation in SKILL.md includes a verification protocol that recommends running shell commands via subprocess.run to call rsvg-convert, allowing the agent to render and compare output images.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes image files and bounding box metadata provided by the agent. While these are handled via structured libraries (OpenCV, NumPy), they represent a data ingestion surface that could theoretically be targeted by malicious inputs.\n
- Ingestion points: External image files (image_path) and agent-provided bounding boxes (focus_targets, focus_edges) are ingested by portrait_mode.py.\n
- Boundary markers: None implemented for the data ingestion logic.\n
- Capability inventory: The skill has file system read access (cv2.imread), network download capability (urllib.request.urlretrieve), and the workflow suggests command execution via subprocess.run.\n
- Sanitization: The skill relies on standard computer vision libraries to process raw image data and bounding box coordinates.
Audit Metadata