The Agent Skills Directory

[SAFE]: The skill correctly instructs users to manage sensitive credentials like ULTRALYTICS_API_KEY through environment variables rather than hardcoding them into scripts or instructions.
[SAFE]: Software dependencies are limited to the official ultralytics package and related CLI tools (yolo, ul) provided by the vendor. The installation instructions (pip install -U ultralytics) target a well-known package from a trusted registry.
[SAFE]: Network operations are scoped to the official Ultralytics Platform infrastructure (platform.ultralytics.com, ultralytics.com) for tasks such as model training, metrics streaming, and weight downloads.
[INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface via the source argument in yolo predict and yolo track, which accepts external URLs. While this could theoretically be used to ingest malicious visual data designed to influence the agent's interpretation of results, this is a standard feature of computer vision tools and no specific injection vulnerabilities were detected.
Ingestion points: source parameter in CLI/Python predict and track modes (e.g., SKILL.md).
Boundary markers: Not explicitly defined in instructions.
Capability inventory: CLI execution of vision tasks, file system writes for results (runs/ directory), and network communication with the Ultralytics Platform.
Sanitization: Standard vision model preprocessing; no specific text-based sanitization for agent interpolation is mentioned.

yolo