ai-ml-development
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill demonstrates Retrieval-Augmented Generation (RAG) and LLM evaluation patterns which inherently involve processing untrusted data. \n- Ingestion points: The RAG implementation in SKILL.md ingests external documents through RecursiveCharacterTextSplitter and Chroma vector stores. \n- Boundary markers: Prompt templates use standard delimiters (e.g., Context: {context}) to separate data from instructions. \n- Capability inventory: The skill uses network tools (ChatOpenAI, ChatAnthropic) and local file storage (Chroma.from_documents). \n- Sanitization: Standard LangChain and LlamaIndex interpolation methods are used, which is appropriate for a development guide. \n- [DYNAMIC_EXECUTION]: The model serving example uses torch.load to deserialize a model file. \n- Evidence: The FastAPI snippet in SKILL.md contains model = torch.load("model.pt"). \n- Analysis: While torch.load uses the pickle module and can be a vector for arbitrary code execution if the model file is untrusted, its use here with a local file path is standard practice for ML deployment documentation. \n- [EXTERNAL_DOWNLOADS]: The skill references models and datasets from well-known hubs. \n- Evidence: References to meta-llama/Llama-2-7b-hf and Hugging Face's imdb dataset. \n- Analysis: These are well-known and trusted repositories in the AI community; their inclusion is standard for ML development workflows.
Audit Metadata