paidf-curation-and-retrieval

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill orchestrates the execution of NVIDIA Cosmos Curator curation and TAO Data Services mining pipelines via make targets and docker commands. These commands are necessary for its documented functionality, such as make run-pipeline and make run_image_pipeline.
  • [EXTERNAL_DOWNLOADS]: The skill facilitates downloading Docker images from the NVIDIA Container Registry (nvcr.io), model weights from HuggingFace, and utility software from conda-forge. These are recognized as trusted vendor resources and well-known services.
  • [CREDENTIALS_SAFE]: The skill includes clear security guidelines for handling sensitive environment variables, such as AWS_SECRET_ACCESS_KEY and various API keys. It instructs the user to use secure secret managers and runtime injection while explicitly warning against hardcoding or logging secret values.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external video and image data to generate AI-derived captions, classification labels, and metadata.
  • Ingestion points: Data is ingested from user-defined local or S3 paths specified in configuration files (e.g., input_video_path).
  • Boundary markers: The skill implements an 8-element VLM prompt framework that includes "visual-grounding constraints" to prevent models from obeying potential instructions hidden in the visual data.
  • Capability inventory: The tool chain has capabilities to write files to the local system or S3 and execute containerized curation workflows.
  • Sanitization: The workflow incorporates multi-stage quality filters (aesthetic, motion, and semantic) and human-in-the-loop review phases (such as the calibration interview) to validate outputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 10:34 PM
Security Audit — agent-trust-hub — paidf-curation-and-retrieval