tao-port-huggingface-model
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads model configurations, weights, and metadata from HuggingFace (
huggingface.co). It also pulls Docker images from NVIDIA's NGC registry (nvcr.io) and clones TAO Toolkit source code repositories from NVIDIA's GitLab/GitHub instances. These operations target trusted organizations and well-known services. - [COMMAND_EXECUTION]: Extensive use of shell commands is prescribed for repository management, Docker container orchestration, and Python package installation. All execution is scoped to the local working directory and specified Docker environments.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of HuggingFace model identifiers and their associated metadata. This data influences subsequent code generation steps. The skill includes a dedicated inspection phase (Phase 1) to validate the model type and task before proceeding, which mitigates risks associated with malicious external model data.
- [PRIVILEGE_ESCALATION]: Documentation notes the potential need for
sudo rm -rfto clean up build artifacts (e.g.,__pycache__,.egg-info) created by root-owned processes inside Docker containers. This is a standard administrative task for managing file permissions in containerized development environments.
Audit Metadata