jetson-video-pipeline
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes external binaries and scripts via subprocess calls in scripts/_pipeline_provenance.py and scripts/_pipeline_runtime.py. These calls are used to invoke NVIDIA SDK samples, system utilities like dpkg and nvcc, and curl for file retrieval. The implementation correctly uses shell=False and validates command arguments.
- [EXTERNAL_DOWNLOADS]: In scripts/pipeline_controller.py, the _retrieve function uses curl to download media content from a user-supplied URL. The skill validates the URL to ensure it uses HTTP/HTTPS protocols and does not contain embedded credentials.
- [DYNAMIC_EXECUTION]: The skill employs dynamic loading and execution for probing local environments. In scripts/pipeline_controller.py, importlib.import_module is used to load the local encode_controller script. Additionally, scripts/_pipeline_provenance.py defines hardcoded Python script templates that are executed via python -c to gather environment details.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, creating a surface for indirect prompt injection.
- Ingestion points: Media metadata such as license and attribution strings, and user-provided URLs/paths in scripts/pipeline_controller.py.
- Boundary markers: The skill requires strict artifact hashing and canonicalization of paths before use.
- Capability inventory: The skill can execute subprocesses, perform network downloads, and write files to the workspace.
- Sanitization: URLs are validated; however, the skill does not explicitly filter metadata strings for potential LLM instructions before they are processed by the agent.
Audit Metadata