The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external machine learning models and image datasets, which constitutes an ingestion surface for untrusted data that could potentially be used for adversarial attacks.
Ingestion points: The skill instructions involve loading external model weights (e.g., arcface_r100.pth in SKILL.md) and large sets of calibration images (load_representative_faces).
Boundary markers: Absent. The instructions do not include specific delimiters or validation steps to ensure the integrity of the imported models or data before processing.
Capability inventory: The skill utilizes shell command execution via trtexec for model compilation and file system write operations to save optimized engines (.plan files).
Sanitization: No sanitization or verification procedures (such as hash checks for models or format validation for images) are described in the instructions.
[SAFE]: The primary purpose and implementation of the skill are consistent with standard professional machine learning engineering practices. The code snippets correctly utilize authorized tools such as NVIDIA TensorRT, PyTorch, and ONNX to achieve legitimate performance optimizations. No malicious patterns such as credential exfiltration, persistence, or obfuscation were detected.

tensorrt