foundation-models-on-device
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents patterns for passing user-provided strings directly into language model sessions without demonstrating explicit sanitization or boundary markers.
- Ingestion points: The
LanguageModelSession.respond(to:)andstreamResponse(to:)methods described inSKILL.mdingest arbitrary user-controlled prompt strings. - Boundary markers: The provided code snippets do not implement delimiters or specific instructions to isolate user content from the session's core instructions.
- Capability inventory: The skill patterns include a
Toolprotocol that allows the model to invoke custom functions, such asRecipeSearchTool, based on prompt processing results. - Sanitization: There is no demonstration of input validation, filtering, or escaping logic to mitigate potential prompt injection within the provided integration patterns.
Audit Metadata