The Agent Skills Directory

[SAFE]: The skill provides detailed guidance on securing the model lifecycle, including the use of immutable revision hashes for Hugging Face checkpoints and Docker container digests for NIM deployments to prevent supply chain attacks.
[SAFE]: The included Python script for interacting with the NIM API follows security best practices: it uses local loopback addresses (127.0.0.1) for microservice communication, implements a dry-run mode, validates request/response integrity using SHA-256 digests, and performs atomic file writes with strict permissions (0o600).
[SAFE]: The skill explicitly warns against dangerous practices such as enabling trust_remote_code without review and loading unverified pickle-based .pt files, recommending safetensors as a secure alternative.
[COMMAND_EXECUTION]: The skill documents how to interact with local services via HTTP POST requests and perform file system operations to store generated video artifacts, which are standard functional requirements for this use case.
[EXTERNAL_DOWNLOADS]: The instructions refer to downloading models and containers from well-known and trusted platforms (Hugging Face, NVIDIA NGC) and provide specific guidance on how to perform these downloads securely using offline caches and verified hashes.

nvidia-cosmos-video