privatelm-cross-platform-llm-client
Pass
Audited by Gen Agent Trust Hub on May 19, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The
ModelController.downloadModelmethod inSKILL.mdimplements functionality to download model files from remote URLs to the device's local storage. This is a standard feature for applications supporting local LLM inference.\n- [PROMPT_INJECTION]: The framework inSKILL.md(specificallyChatControllerandVisionService) creates an indirect prompt injection surface by interpolating user input and images into LLM chat templates.\n - Ingestion points: Untrusted user input enters the agent context through the
sendMessageandprocessImageLocalmethods inSKILL.md.\n - Boundary markers: Present; the
AppConfig.chatTemplatesclass defines clear delimiters (e.g., ChatML, Llama3) to distinguish between system, user, and assistant roles.\n - Capability inventory: The skill involves network operations to cloud API providers, file writes for model storage, and persistent storage using Hive.\n
- Sanitization: No explicit sanitization of user-provided prompt content is shown in the snippets, which is expected behavior for a client-side library framework.
Audit Metadata