deepstream-import-vision-model

Pass

Audited by Gen Agent Trust Hub on Aug 18, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: Instructions in SKILL.md direct the agent to 'Run the full pipeline autonomously without pausing for confirmation at each step', which reduces human oversight during high-privilege operations like package installation and file system modification.
  • [EXTERNAL_DOWNLOADS]: The skill fetches model artifacts, configurations, and dependencies from well-known external domains including huggingface.co and api.ngc.nvidia.com.
  • [COMMAND_EXECUTION]: The workflow performs automated C++ compilation using g++ for custom bounding box parsers and utilizes destructive commands such as 'rm -rf' in cleanup scripts, albeit with input validation.
  • [REMOTE_CODE_EXECUTION]: The pre-flight and model acquire phases install system-level packages via apt-get and Python libraries via pip from external repositories to establish the TensorRT build environment.
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting and processing untrusted metadata from external repository configurations. 1. Ingestion points: config.json and HuggingFace API response data. 2. Boundary markers: Absent for metadata processing steps. 3. Capability inventory: C++ compilation, package installation, and directory deletion. 4. Sanitization: Includes input validation for model identifiers and path traversal checks for report image embedding.
  • [COMMAND_EXECUTION]: Dynamically generates and executes Makefiles and Python scripts at runtime to handle model-specific label extraction, parser build logic, and benchmark data processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 18, 2026, 01:39 AM
Security Audit — agent-trust-hub — deepstream-import-vision-model