The Agent Skills Directory

[Indirect Prompt Injection] (MEDIUM): The skill documents code patterns for Query Expansion and HyDE (Hypothetical Document Embeddings) that are vulnerable to prompt injection.
Ingestion points: The query parameter in the expand_query and hyde functions in SKILL.md takes untrusted user input.
Boundary markers: None. The snippets use raw f-strings (f"...\n{query}") which provide no separation between instructions and data.
Capability inventory: The snippets call model.generate(prompt), which influences the LLM's logic and subsequent retrieval steps.
Sanitization: No sanitization, escaping, or validation of the query input is present.
[EXTERNAL_DOWNLOADS] (LOW): The code snippets reference external Python libraries (rank_bm25, sentence_transformers, faiss) which are standard in the industry but are listed without version constraints or integrity checks.

rag-systems