skills/nvidia/skills/deepstream-sop/Gen Agent Trust Hub

deepstream-sop

Pass

Audited by Gen Agent Trust Hub on Jul 2, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill's Docker configuration includes instructions to download the Basler Pylon SDK from an Azure-hosted CDN and to clone GStreamer plugins from Basler's official GitHub organization. It also references the DDM research repository on GitHub for core model functionality. These downloads originate from well-known vendors and established research groups.
  • [COMMAND_EXECUTION]: The service utilizes subprocess.run and subprocess.check_output to execute system-level commands such as nvidia-smi for performance monitoring and an internal Python script for TensorRT engine optimization. These executions use controlled arguments and are integral to the application's performance tuning and monitoring workflows.
  • [DATA_EXFILTRATION]: The microservice architecture involves processing untrusted media data from files and RTSP streams via vision-language models. Results are published to a Kafka messaging topic. While this creates a surface for indirect prompt injection, it is the primary intended function of the service and is implemented with standard isolation patterns.
  • [SAFE]: The skill incorporates robust security practices, including the use of a non-privileged user account within the containerized environment. It also employs established environment variable patterns for handling configuration and sensitive parameters, avoiding hardcoded credentials.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 2, 2026, 12:54 AM
Security Audit — agent-trust-hub — deepstream-sop