tensorrt-optimization
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of custom TensorRT plugins and CUDA kernels. It provides C++ templates intended to be compiled and integrated into the inference engine at runtime.
- [INDIRECT_PROMPT_INJECTION]: The skill accepts external model files (ONNX, engines) and calibration datasets as input for optimization processes.
- Ingestion points: Model files are read using
parser.parse()and calibration data is consumed vianext(self.data_loader)inSKILL.md. - Boundary markers: No specific delimiters or security warnings are used when processing these external data sources to prevent potential instruction injection hidden in model metadata.
- Capability inventory: The skill utilizes file system access (Read/Write), subprocess execution (
trtexec), and GPU memory management (cuda.mem_alloc) across its instructional examples. - Sanitization: No validation or sanitization of model metadata or calibration data content is performed before processing.
Audit Metadata