tao-train-rtdetr

Pass

Audited by Gen Agent Trust Hub on Jun 18, 2026

Risk Level: SAFE
Full Analysis
  • [METADATA_POISONING]: The skill frontmatter identifies the author as NVIDIA Corporation, which is consistent with the skill's purpose but differs from the provided author context (promptingcompany).
  • [DATA_EXPOSURE_AND_EXFILTRATION]: Telemetry via Weights & Biases (WandB) is enabled by default in the training schemas (e.g., distill.schema.json). WandB is a well-known service for machine learning tracking, and its use here is consistent with standard ML workflows.
  • [INDIRECT_PROMPT_INJECTION]: The skill defines ingestion points for external COCO dataset files, which is essential for training.
  • Ingestion points: annotations.json and label_map.txt (references/skill_info.yaml).
  • Boundary markers: No specific delimiters or instructions to ignore embedded instructions in data are provided.
  • Capability inventory: Execution of rtdetr shell commands via TAO Toolkit (references/skill_info.yaml).
  • Sanitization: No explicit sanitization for dataset content is documented.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 18, 2026, 04:38 PM
Security Audit — agent-trust-hub — tao-train-rtdetr