tao-train-dino
Pass
Audited by Gen Agent Trust Hub on Jun 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill provides command templates and configuration schemas to guide the agent in executing shell commands for model training and deployment using Docker and the
dinoCLI tool. Evidence is found inreferences/tao-deploy-dino.mdandreferences/skill_info.yaml. - [EXTERNAL_DOWNLOADS]: The skill references datasets and container images hosted on NVIDIA-controlled infrastructure. Evidence includes references to
s3://nvcf-storage-handling/for training data andnvcr.io/nvidia/tao/for tool containers inSKILL.mdandreferences/tao-deploy-dino.md. - [PROMPT_INJECTION]: The skill specifies ingestion of external data files (COCO JSON) which presents a surface for indirect prompt injection, though this is typical for machine learning training tasks. Ingestion points: dataset URIs defined in
SKILL.mdandreferences/dino-data-specs.md. Boundary markers: None explicitly defined for file contents. Capability inventory: execution of thedinoCLI anddocker runcommands as described inreferences/skill_info.yaml. Sanitization: relies on the internal parsing logic of the TAO Toolkit.
Audit Metadata