The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media files for object detection and tracking, which presents a surface for potential instructions embedded in file metadata or visual content to influence agent behavior.
Ingestion points: scripts/video_analyzer.py reads video frames via cv2.VideoCapture.
Boundary markers: There are no explicit boundary markers or instruction filtering applied to the content of processed frames.
Capability inventory: The skill possesses file-writing capabilities and model-training logic.
Sanitization: Input media files are not sanitized for malicious metadata or adversarial visual patterns targeting LLM instructions.
[EXTERNAL_DOWNLOADS]: Reference documentation suggests installing external dependencies from the internet.
The references/yolo-guide.md file suggests installing the ultralytics package via pip.
The references/tracking-algorithms.md file suggests cloning the abewley/sort repository from GitHub to implement tracking.
[DYNAMIC_EXECUTION]: The model training and inference scripts load PyTorch weight files (.pt), which utilize the pickle module for serialization. Loading weights from untrusted sources could lead to arbitrary code execution due to the nature of the pickle format, though this is a standard requirement for the skill's primary purpose.

computer-vision-pipeline