local-ai-models
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs users to download the MLX framework and related examples from the official Apple-maintained GitHub organization (ml-explore). It also references models from Hugging Face Hub, which is a standard repository for machine learning assets.
- [COMMAND_EXECUTION]: The model quantization guide provides shell commands for model conversion using the 'mlx-lm' package and CLI tools. These instructions are documented for the developer's local environment and do not involve silent or malicious execution.
- [INDIRECT_PROMPT_INJECTION]: The skill defines patterns for processing user input and images in local LLM applications. While it describes attack surfaces common to AI apps (such as tool calling and data extraction), it does so within the context of a developer guide.
- Ingestion points: User-provided text in chat interfaces and image data in Vision Language Models (VLMs) as described in
references/mlx-swift/advanced-patterns.mdandreferences/mlx-swift/vision-patterns.md. - Boundary markers: The provided prompt templates use standard line breaks and headers but lack specialized sanitization logic in the code snippets.
- Capability inventory: Patterns for tool execution and structured data generation are provided in the reference code.
- Sanitization: Code snippets focus on implementation logic and do not explicitly demonstrate prompt sanitization or advanced injection mitigation.
Audit Metadata