The Agent Skills Directory

Indirect Prompt Injection (LOW): The guides describe patterns for ingesting untrusted data from web pages, GitHub repositories, and local directories. This is a standard RAG surface where malicious instructions in documents could attempt to influence agent behavior.
Ingestion points: SimpleWebPageReader, GithubRepositoryReader, and SimpleDirectoryReader in data_connectors.md.
Boundary markers: Examples in query_engines.md show the use of dashed line delimiters in PromptTemplate to separate context from instructions.
Capability inventory: The documentation shows agents capable of text synthesis and calling localized tools (e.g., multiply in agents.md).
Sanitization: Explicit sanitization or filtering of retrieved content is not implemented in these basic reference snippets.
Data Exposure (SAFE): The documentation uses safe placeholders like integration_token="your-token" for API credentials rather than hardcoding actual secrets.

llamaindex