The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The documentation identifies the risk of indirect prompt injection within RAG (Retrieval-Augmented Generation) architectures.
Ingestion points: The skill describes systems that retrieve external documents from indices which are processed by an LLM (references/07-edge-case-y.md).
Boundary markers: The guidance suggests using clear system boundaries and contracts to manage data flow and trust.
Capability inventory: As an architectural guide, this skill does not contain executable code, but it provides instructions for systems that interact with LLM providers and local runtimes.
Sanitization: The skill explicitly recommends validating model outputs and treating retrieved documents as untrusted inputs (references/01-overview.md).
[PROMPT_INJECTION]: The skill contains text patterns often associated with prompt injection, which were evaluated as benign in context.
Evidence: The file references/07-edge-case-y.md mentions the phrase "ignore previous instructions".
Context: This is listed as a potential security edge case for RAG systems (e.g., a document in the index containing instructions to bypass the system prompt). It is an educational description of a threat rather than an active attempt to compromise the agent.
[EXTERNAL_DOWNLOADS]: The documentation references downloading machine learning models and configurations from external sources.
Evidence: references/01-overview.md discusses using the hf-hub library to fetch weights and configurations from Hugging Face.
Note: References to Hugging Face represent standard integration with a well-known service in the machine learning ecosystem and do not constitute a security risk in this context.

rust-ml-llm-architecture