tao-finetune-video-clip

Pass

Audited by Gen Agent Trust Hub on Sep 23, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes video metadata in JSON format (vadr1_chunks) for training and inference, which constitutes a surface for processing untrusted data.
  • Ingestion points: Files specified in dataset.train.video_text.metadata and dataset.val.video_text.metadata (SKILL.md, references/spec_template_train.yaml).
  • Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded commands within the input metadata.
  • Capability inventory: The skill executes training and inference actions via Docker containers.
  • Sanitization: No explicit validation or filtering of the textual content in the metadata JSON is provided.
  • [EXTERNAL_DOWNLOADS]: Fetches model weights and training datasets from vendor-controlled infrastructure.
  • Uses s5cmd to mirror assets from a vendor internal S3 endpoint (pdx.s8k.io).
  • Pulls container images from the vendor's official Container Registry (nvcr.io).
  • [DYNAMIC_EXECUTION]: The skill instructs the agent to set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1, which disables PyTorch's weights_only unpickling security check.
  • Evidence: This configuration is used in the DOCKER_COMMON environment variables in SKILL.md and referenced in eval.config.
  • Context: This is documented as a requirement for compatibility with specific model checkpoints but reduces protection against malicious pickle data if non-vendor models are loaded.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 23, 2026, 02:11 PM
Security Audit — agent-trust-hub — tao-finetune-video-clip