The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes config.json and preprocessor_config.json files downloaded from external repositories (HuggingFace and NVIDIA NGC). These files are used to extract model labels and architecture parameters that influence code generation (C++ parser) and shell command construction. While the ingestion is performed using standard JSON parsers, the lack of explicit boundary markers or sanitization for the extracted strings presents a potential vulnerability surface.
Ingestion points: config.json and preprocessor_config.json downloaded in scripts/model/hf-download-config.sh and scripts/model/ngc-download.sh.
Boundary markers: Absent.
Capability inventory: The skill has extensive capabilities including shell command execution (subprocess.run), Docker management (docker run), file system operations (rm -rf), and network access (curl, wget).
Sanitization: Content is parsed via json.loads in scripts/model/config-to-labels.py and scripts/model/safetensors_to_onnx.py before being utilized in the pipeline.
[COMMAND_EXECUTION]: The skill relies on executing various shell and PowerShell scripts to orchestrate the model deployment process. This includes building TensorRT engines, compiling C++ parsers, and running performance benchmarks. The skill attempts to mitigate risk by executing compute-heavy and complex tasks within a Docker container environment.
[EXTERNAL_DOWNLOADS]: The skill fetches model weights and configuration files from HuggingFace and NVIDIA NGC. These are recognized as well-known and trusted services. The acquisition scripts (hf-list-files.sh, ngc-download.sh) include validation logic to ensure organization and model names adhere to expected character sets.
[PRIVILEGE_ESCALATION]: The report generation script (scripts/report/md-to-html-pdf.py) contains logic to automatically install the wkhtmltopdf system package using sudo apt-get install if the binary is missing. This operation is documented as part of the environment setup process, typically intended to run within the isolated DeepStream container.

deepstream-import-vision-model