The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an architecture that processes external data and models, creating a potential surface for indirect prompt injection.
Ingestion points: The skill describes loading machine learning models from external URLs or servers and performing inference on untrusted input data (images, audio, and text) using MLModel.prediction(from:).
Boundary markers: The provided code patterns do not include explicit delimiters or instructions to help the agent distinguish between data and embedded instructions within the processed inputs.
Capability inventory: The skill utilizes file-system write access through the ModelManager class to cache compiled models in the Application Support directory, and it leverages the device's compute units (CPU, GPU, and Neural Engine).
Sanitization: No sanitization or signature verification of external model packages is demonstrated before they are compiled and executed by the system.
[DYNAMIC_EXECUTION]: The skill utilizes runtime compilation and loading features of the Core ML framework.
Evidence: The code employs MLModel.compileModel(at:) to transform .mlpackage source files into executable .mlmodelc formats on-device and uses MLModel(contentsOf:) to load these compiled binaries into the application process.
[SAFE]: The skill adheres to modern Swift security and performance standards.
Best Practices: It emphasizes asynchronous loading (await MLModel.load) to maintain application responsiveness and uses actor-based caching to ensure thread safety and prevent resource contention during model operations.
Standard APIs: All described functionality uses official Apple frameworks (Core ML, Vision, Natural Language) according to their intended design.

coreml