The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external and potentially untrusted data through machine learning models, creating a surface for indirect prompt injection or adversarial attacks.\n
Ingestion points: The skill ingests images via createPixelBuffer and text via analyzeSentiment in SKILL.md and references/coreml-swift-integration.md.\n
Boundary markers: Absent. Data inputs are passed directly to the MLModel.prediction API without delimiters or specific instructions to ignore embedded content.\n
Capability inventory: The skill utilizes MLModel.prediction for inference and FileManager.default.moveItem for persistent model storage in references/coreml-swift-integration.md.\n
Sanitization: Absent. The provided snippets do not include logic to sanitize or validate inputs against malicious patterns or adversarial examples.\n- [DYNAMIC_EXECUTION]: The skill provides implementation patterns for on-device compilation of model source files into executable binaries.\n
The ModelManager implementation in references/coreml-swift-integration.md uses MLModel.compileModel(at:) to transform .mlpackage or .mlmodel files into the .mlmodelc format at runtime.\n
The instructions in SKILL.md explicitly note this is intended for models "downloaded from a server," which involves executing assets generated from remote sources on the local device.\n- [METADATA_POISONING]: The skill contains deceptive documentation references in the metadata and references section.\n
The References section in SKILL.md contains multiple links using the sosumi.ai domain instead of the official developer.apple.com domain. While this is a known community redirector, it imitates official Apple documentation paths and directs users to unofficial sources for platform information.

coreml