The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements RAG patterns that ingest untrusted external data (retrieved context) and user-supplied queries, which are then interpolated directly into prompts. This creates an attack surface where malicious instructions in the documents could influence the agent's behavior.
Ingestion points: External data enters the system context via the context and documents variables in examples/basic_rag.py and examples/agentic_rag.py.
Boundary markers: The prompt templates in the examples (e.g., generate_answer and evaluate_retrieved_docs) do not utilize delimiters or specific instructions to disregard embedded commands in the context data.
Capability inventory: The system leverages the untrusted context to perform generation tasks and agentic reasoning decisions, such as those found in the AgenticRAG.decide_retrieval method.
Sanitization: The implementation lacks evidence of text sanitization, validation, or escaping for the external content before it is processed by the model.

rag-agent-builder