The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/inference_optimizer.py uses an unsafe deserialization method to analyze PyTorch model files.
Evidence: Lines 145 and 239 in scripts/inference_optimizer.py call torch.load(str(self.model_path), map_location='cpu') and torch.load(str(self.model_path), map_location=device) respectively.
Risk: By default, torch.load uses the Python pickle module, which is vulnerable to arbitrary code execution during deserialization. If an attacker provides a malicious .pt or .pth model file and a user attempts to benchmark or analyze it with this script, it could result in full system compromise. Using weights_only=True is the recommended mitigation, which is not implemented here.
[INDIRECT_PROMPT_INJECTION]: The skill's scripts process untrusted external data (datasets, annotations, and model structures) which provides an attack surface for indirect prompt injection.
Ingestion points: scripts/dataset_pipeline_builder.py reads COCO JSON, Pascal VOC XML, and YOLO TXT annotation files; scripts/vision_model_trainer.py reads JSON and YAML configuration files; scripts/inference_optimizer.py reads model files (.pt, .onnx).
Boundary markers: Absent. There are no explicit instructions to ignore embedded commands or markers within the processed data.
Capability inventory: The scripts have capabilities for file system writes (json.dump, yaml.dump, Path.write_text), directory creation, and potential code execution through unsafe model loading.
Sanitization: Inconsistent. While vision_model_trainer.py uses yaml.safe_load(), other scripts process JSON and XML (via xml.etree.ElementTree) without robust sanitization of the content against malicious patterns.

senior-computer-vision