router-builder
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill utilizes external dependencies and downloads a pre-trained model from a well-known service.\n
- Evidence: The skill requires
semantic-router,sentence-transformers, andpyyaml. It also utilizes theHuggingFaceEncoderto download theall-MiniLM-L6-v2model from Hugging Face.\n- [INDIRECT_PROMPT_INJECTION]: The skill exposes a routing tool that processes user-supplied text to determine the next task or agent. This input represents a potential vector for indirect prompt injection.\n - Ingestion points: The
route_querytool inmcp/servers/router-server/server.pyingests a query string directly from the agent's interaction context.\n - Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands within the input query.\n
- Capability inventory: The skill reads local YAML configuration files and uses an embedding-based classification engine to select resources.\n
- Sanitization: The input query is not sanitized or filtered before being passed to the semantic classification engine.
Audit Metadata