The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the implementation of Retrieval-Augmented Generation (RAG) pipelines and Agentic Workflows. These patterns are inherently susceptible to indirect prompt injection if the data being retrieved (e.g., from vector stores or web searches) contains malicious instructions designed to manipulate the LLM's behavior.
Ingestion points: Data retrieved from FAISS vector stores and tool outputs (e.g., calculate_shipping) are interpolated into prompts.
Boundary markers: The RAG example uses a template: "Answer the question based only on context: {context}\nQuestion: {question}". While this provides some instruction, it lacks robust delimiter-based isolation for the untrusted context.
Capability inventory: The skill showcases tool binding (bind_tools), LCEL execution, and LangGraph state management, which could be misused if an injection successfully hijacks the agent's flow.
Sanitization: The skill does not explicitly demonstrate or mandate input sanitization or verification of retrieved content before prompt interpolation.

langchain