The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes torchvision.models with pretrained=True in SKILL.md to initialize model architectures. This triggers the download of pretrained weights from PyTorch's official repository, which is a standard and expected operation for computer vision tasks.
[REMOTE_CODE_EXECUTION]: The DefectDetectionModel class in SKILL.md uses torch.load(model_path) to load model state dictionaries. This function uses the pickle module for deserialization by default, which can be exploited to execute arbitrary code if the model_path points to a malicious file. This represents a potential path for code execution via untrusted model files.
[PROMPT_INJECTION]: The skill is designed to process external images and structured data (CSV, Excel, JSON) as described in instructions.md and SKILL.md, establishing an indirect prompt injection surface.
Ingestion points: Untrusted data enters via image_path parameters in the predict, detect, and perform_inspection methods in SKILL.md, as well as via file imports described in instructions.md.
Boundary markers: Absent. The skill does not implement delimiters or safety instructions to distinguish between data and control commands when processing external files.
Capability inventory: The skill possesses filesystem write capabilities via pd.ExcelWriter and potential for code execution through unsafe model loading in SKILL.md.
Sanitization: Absent. There is no evidence of validation or sanitization of content extracted from processed files before it enters the execution context.

defect-detection-ai