foundation-models-on-device

Pass

Audited by Gen Agent Trust Hub on Feb 24, 2026

Risk Level: SAFE
Full Analysis
  • [DATA_EXFILTRATION]: The skill emphasizes on-device processing using Apple's SystemLanguageModel, which is designed to prevent data from leaving the local device, enhancing privacy.
  • [COMMAND_EXECUTION]: While the skill includes patterns for 'Tool Calling', these tools execute custom Swift code within the app's own sandbox for domain-specific tasks and do not represent arbitrary command execution risks in the context of the skill itself.
  • [INDIRECT_PROMPT_INJECTION]: The skill describes an architecture that processes user-provided natural language to trigger tools or generate structured data.
  • Ingestion points: User input is passed to session.respond(to:) and parameters are parsed into @Generable structs.
  • Boundary markers: The documentation specifically recommends using system instructions to implement safety measures (e.g., 'Respond with I can't help with that for dangerous requests').
  • Capability inventory: The system can trigger custom Tool implementations through func call(arguments:) based on model interpretation of input.
  • Sanitization: The skill promotes the use of @Generable structured output to reduce the risks associated with parsing raw strings, though input sanitization remains the responsibility of the implementing developer.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 24, 2026, 07:52 PM