vector-search-patterns
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
rag_queryfunction inSKILL.mddemonstrates a retrieval-augmented generation pipeline that contains an indirect prompt injection vulnerability surface. - Ingestion points: The skill ingests untrusted data from both the user-supplied
questionand external content retrieved from a database via thesearchfunction. - Boundary markers: The prompt template uses simple text labels ('Context:', 'Question:') to delimit content but lacks robust separators or instructions for the model to ignore commands potentially embedded within the retrieved context.
- Capability inventory: The skill uses network and database capabilities; the primary risk occurs when the combined context is processed by
llm.complete, which could influence the agent's logic if it has access to sensitive tools. - Sanitization: There is no evidence of sanitization, escaping, or filtering of the retrieved chunks before they are incorporated into the prompt.
- [SAFE]: The skill uses
httpxto generate embeddings via OpenAI's official API (https://api.openai.com/v1/embeddings). As this involves communication with a well-known service for its standard purpose, this network behavior is considered safe.
Audit Metadata