tao-train-nvpanoptix3d
Pass
Audited by Gen Agent Trust Hub on Aug 25, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill invokes the TAO Toolkit via a specific Python module entrypoint: 'python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d'. This is the standard method for executing training, evaluation, and inference tasks within the toolkit image.
- [INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it ingests untrusted data from external dataset files.
- Ingestion points: The agent reads metadata from JSON files specified by 'dataset.train.json_path', 'dataset.val.json_path', and 'dataset.test.json_path' in 'references/skill_info.yaml'.
- Boundary markers: The instructions do not define delimiters or specific warnings to ignore instructions that might be embedded within these metadata files.
- Capability inventory: The skill has the capability to execute shell commands and write to the filesystem via the TAO entrypoint defined in 'references/skill_info.yaml'.
- Sanitization: There is no evidence of content sanitization or validation for the ingested JSON data in the skill's instructions or schemas.
- [SAFE]: No hardcoded credentials, malicious obfuscation, or unauthorized network exfiltration patterns were detected. The use of 'encryption_key' parameters is documented as a standard tool configuration for model security.
Audit Metadata