deepstream-import-vision-model

Pass

Audited by Gen Agent Trust Hub on Jul 15, 2026

Risk Level: SAFE
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill downloads model weights, configurations, and metadata from HuggingFace and NVIDIA NGC (catalog.ngc.nvidia.com). These are well-known technology services and trusted vendor repositories.
  • [DATA_EXFILTRATION]: The report generation workflow implements specific security measures to prevent data exfiltration:
    • The md-to-html-pdf.py script explicitly disables local file access in the wkhtmltopdf engine.
    • It contains a path traversal check to ensure images embedded in reports originate only from the expected model directory.
    • cleanup.sh uses regex validation (^[A-Za-z0-9._-]+$) and explicitly rejects parent directory segments (..) to prevent accidental or malicious file deletion outside the project scope.
  • [COMMAND_EXECUTION]: Shell commands involving model names and paths are sanitized using regex and explicit checks in scripts like hf-download-config.sh and ngc-download.sh to prevent command injection.
  • [SAFE]: The skill uses isolated Python virtual environments (build/.venv_optimum) to manage dependencies, preventing interference with the host system's global packages.
  • [SAFE]: The Mermaid diagram renderer (render-mermaid-for-pdf.py) includes a whitelist for Puppeteer configuration files, ensuring that only vetted, local configurations can be used to control the headless browser, mitigating risks associated with browser-based attacks.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 15, 2026, 05:41 AM
Security Audit — agent-trust-hub — deepstream-import-vision-model