local-rag-search

Warn

Audited by Gen Agent Trust Hub on Feb 14, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • EXTERNAL_DOWNLOADS (MEDIUM): The README instructions recommend installing the mcp-local-rag prerequisite server from an untrusted GitHub repository (nkapila6/mcp-local-rag) using the uvx command. This pattern involves downloading and executing code from an unverified external source during environment setup.
  • PROMPT_INJECTION (LOW): The skill provides a high surface area for Indirect Prompt Injection (Category 8) as it is designed to ingest and process arbitrary web content. Evidence chain:
  • Ingestion points: Data retrieved from DuckDuckGo, Google, and other search backends via tools like rag_search_google and deep_research in SKILL.md.
  • Boundary markers: Absent; there are no instructions for the agent to use delimiters or to disregard instructions found within the retrieved search results.
  • Capability inventory: Primarily data retrieval and synthesis (LOW tier capability), though it influences agent reasoning and downstream responses.
  • Sanitization: Absent; the skill does not include logic or instructions to filter or escape malicious instructions embedded in web pages.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 14, 2026, 04:26 PM