The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references the official apple/foundation-models-utilities package from GitHub (https://github.com/apple/foundation-models-utilities). This is a legitimate resource from a trusted organization intended to provide emerging patterns for the Foundation Models framework.
[INDIRECT_PROMPT_INJECTION]: The skill implements features that ingest untrusted user data, such as prompts and images, for processing by language and vision models, which presents an attack surface for indirect injection.
Ingestion points: User-supplied content is ingested via the respond(to:) method and Attachment segments in the Foundation Models framework.
Boundary markers: The skill emphasizes the "Instructions vs Prompts" hierarchy to prioritize developer-defined safety rules over potentially malicious user input, as described in foundation-models/safety-and-guardrails.md.
Capability inventory: The skill supports tool calling (executing Swift code), network operations via Private Cloud Compute or third-party backends, and reading files from ~/Downloads/docs/ for extra documentation context.
Sanitization: The skill utilizes the framework's built-in safety stack, including input/output guardrails and instruction-based constraints, to filter harmful generated content.

apple-intelligence