The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill pulls the official TAO Toolkit container from the NVIDIA container registry (nvcr.io) and fetches pretrained model components from HuggingFace (nvidia/Cosmos-Embed1-224p). These are verified vendor resources essential for the skill's stated purpose.- [COMMAND_EXECUTION]: Model operations such as training, evaluation, and inference are executed via the cosmos-embed1 CLI inside a Docker container. These commands are static and follow security best practices for resource isolation and GPU management.- [DYNAMIC_EXECUTION]: A runtime installation of the 'protobuf' library is performed via pip inside the container to ensure environment compatibility. This is a standard maintenance task using verified package registries to resolve known dependency conflicts.- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local JSON metadata files (e.g., msrvtt_test_1k.json) and video files. Boundary markers to delimit untrusted data are absent in the prompt templates. The skill's capabilities include subprocess execution within the container and network access for downloading dependencies. No explicit sanitization of metadata content is described. The risk is evaluated as safe because the data is consumed by specialized embedding models rather than a general-purpose instruction-following agent.- [DATA_EXFILTRATION]: Securely handles HuggingFace authentication tokens using environment variables (-e HF_TOKEN) and persistent cache mounts. This method adheres to security best practices for avoiding credential leakage in logs, specifications, or code.

tao-finetune-cosmos-embed