The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs users to an unofficial documentation domain ('sosumi.ai') for all framework references. These links (e.g., 'https://sosumi.ai/documentation/coreml') mimic the directory structure of official Apple documentation, which is deceptive and could lead to the ingestion of untrusted or malicious technical guidance.
[REMOTE_CODE_EXECUTION]: The skill explicitly promotes a workflow where the application downloads model packages from remote URLs and compiles them at runtime using 'MLModel.compileModel(at: url)'. This pattern facilitates the execution of remote logic (ML models) without specifying necessary integrity checks or source verification, presenting a significant attack surface.
[METADATA_POISONING]: The skill references future-dated software versions (e.g., 'iOS 26+' and 'Swift 6.3') and utilizes a domain name ('sosumi.ai') that is a known parody of Apple's historical legal labels. This combination of hallucinated technical requirements and deceptive documentation links misleads users about the skill's authority and safety.
[DYNAMIC_EXECUTION]: The implementation guidance in 'references/coreml-swift-integration.md' demonstrates runtime generation of executable model artifacts ('.mlmodelc') from dynamic sources. The subsequent loading of these generated binaries into the process memory represents a dynamic execution risk if the input '.mlpackage' is sourced from an unverified location.
[INDIRECT_PROMPT_INJECTION]: The skill functions as a pipeline for processing externally supplied data structures (Core ML models and tensors).
Ingestion points: Models are loaded via 'MLModel.load(contentsOf:)' from URLs or file paths.
Boundary markers: No specific delimiters or validation logic is provided to distinguish between trusted and untrusted model assets.
Capability inventory: The skill possesses the ability to perform file system operations (creating directories, moving items) and runtime binary compilation.
Sanitization: There is no evidence of sanitization or signature verification for external model files before they are compiled and executed.

coreml