databricks-vector-search

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: Indirect Prompt Injection surface detected. The skill facilitates the creation of RAG pipelines that ingest data from external vector indexes and interpolate it into agent prompts, which can be exploited if the source data contains malicious instructions.
  • Ingestion points: Data is ingested from vector indexes using tools such as query_vs_index, scan_vs_index, and the WorkspaceClient.vector_search_indexes.query_index method.
  • Boundary markers: Code examples (specifically in end-to-end-rag.md) use simple natural language prefixes (e.g., "Answer using this context:") rather than robust delimiters or explicit "ignore instructions" warnings for the retrieved content.
  • Capability inventory: The documented agent workflows incorporate high-privilege capabilities including workspace infrastructure management, model serving endpoint interaction, and SQL execution via the execute_sql tool.
  • Sanitization: The provided implementation patterns do not demonstrate explicit sanitization, validation, or filtering of retrieved content before it is processed by the model.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 02:51 PM
Security Audit — agent-trust-hub — databricks-vector-search