The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes patterns for building AI agents that ingest data from external sources, such as databases and retrieved documents. This architectural surface is susceptible to indirect prompt injection, where malicious instructions within the processed data could attempt to subvert the agent's logic.
Ingestion points: The RAG implementation in references/details.md fetches external document content, and the search_database tool in SKILL.md processes external query results.
Boundary markers: While the examples use standard prompt instructions, they do not demonstrate the use of robust delimiters or specific "ignore embedded instructions" blocks to protect against adversarial data.
Capability inventory: The documentation includes examples of tools with external side effects, such as the send_email coroutine in references/details.md, which could be abused if an injection is successful.
Sanitization: The reference implementations do not include explicit validation or sanitization steps for external content prior to its inclusion in LLM prompts.

langchain-architecture