foundation-models-on-device
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides patterns for building applications that process untrusted user input via Apple's on-device LLM session methods.
- Ingestion points: Data enters the model context through
session.respond(to:)andsession.streamResponse(to:)calls demonstrated inSKILL.md. - Boundary markers: The skill suggests using the
instructionsparameter to define the model's role and safety measures, which serves as a defensive instructional frame. - Capability inventory: The skill documents the
Toolprotocol (e.g.,RecipeSearchToolinSKILL.md), which allows the language model to trigger Swift code execution and external data retrieval (e.g.,searchRecipes). - Sanitization: The provided patterns do not include explicit input validation or sanitization logic for the natural language prompts before they are passed to the model.
Audit Metadata