The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes computer vision datasets consisting of images and structured annotations in COCO (JSON), Pascal VOC (XML), and YOLO (TXT) formats. This provides a surface for indirect prompt injection if the agent interprets text content within these files as instructions.
Ingestion points: The scripts/dataset_pipeline_builder.py and scripts/vision_model_trainer.py scripts read images and structured annotation files from local directories.
Boundary markers: The skill does not implement delimiters or explicit instructions for the agent to ignore potentially malicious content embedded within annotation metadata.
Capability inventory: The skill includes scripts for file system management (read, write, symlink) and provides workflows for model training, optimization, and deployment.
Sanitization: Standard libraries are used for parsing (json, xml.etree.ElementTree, yaml), but there is no semantic sanitization of the annotation content.
[DYNAMIC_EXECUTION]: The scripts/inference_optimizer.py script uses the torch.load() function for model analysis.
Evidence: Found in the _analyze_pytorch and _benchmark_pytorch methods of scripts/inference_optimizer.py.
Finding: The torch.load() function utilizes the Python pickle module, which is susceptible to arbitrary code execution if a maliciously crafted model file is loaded. While common in PyTorch workflows, this constitutes unsafe deserialization of local data.
[EXTERNAL_DOWNLOADS]: Fetches configuration and pre-trained weights from well-known frameworks.
Evidence: scripts/vision_model_trainer.py references detectron2:// URIs for official model zoo resources.
Finding: The skill integrates with established computer vision frameworks that may download official pre-trained weights during initial execution.

senior-computer-vision