The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages the augmentation pipeline using Docker and the uv tool. It provides extensive hardening guidance, including the use of isolated network bridges (--network cdf), capability dropping (--cap-drop=ALL), and preventing privilege escalation (no-new-privileges:true).
[CREDENTIALS_SAFE]: Security best practices are followed for credential handling. API keys and cloud credentials are provided via environment variables (referenced by api_key_env) rather than being hardcoded in configuration files or scripts. The documentation explicitly warns against secret exposure in logs or version control.
[EXTERNAL_DOWNLOADS]: The skill communicates with official NVIDIA hosted endpoints and standard cloud storage backends. It uses a specialized storage client to manage remote file access safely without exposing sensitive local paths.
[INDIRECT_PROMPT_INJECTION]: While the skill processes external images and videos that could theoretically contain hidden instructions, it employs structural boundaries (Pydantic-validated YAML) and specific behavioral instructions (e.g., grounding VLMs in 'visible facts' only) to protect the agent's integrity during media analysis.

cdf-augmentation