The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the implementation of retrieval-augmented generation (RAG) systems that ingest data from various external and potentially untrusted sources, including web pages via SimpleWebPageReader, GitHub repositories via GithubRepositoryReader, and external APIs via JSONReader. The provided code snippets demonstrate interpolating this external data directly into prompt templates (e.g., the qa_prompt and refine_prompt examples) without providing guidance on sanitization, filtering, or the use of strict boundary markers to prevent malicious instructions embedded in the data from influencing the agent's behavior.
Ingestion points: SimpleDirectoryReader, SimpleWebPageReader, GithubRepositoryReader, DatabaseReader, and JSONReader as described in SKILL.md and references/data_connectors.md.
Boundary markers: The examples use simple string interpolation (e.g., Context: {context_str}) which does not inherently prevent the LLM from obeying instructions hidden within the context variable.
Capability inventory: The skill identifies capabilities for file system operations (persist), network communication (load_data from URLs), and database connectivity (DatabaseReader).
Sanitization: The provided documentation and code examples do not include explicit steps for sanitizing, validating, or escaping external content before it is processed by the LLM.

llamaindex