rag-retrieval
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines patterns that ingest external data from web search and vector databases, providing a surface for potential indirect prompt injection attacks through content processed at runtime.\n
- Ingestion points: Document retrieval in
scripts/scripts/crag-workflow.py(via web search) andscripts/rag-pipeline-template.ts(via vector database).\n - Boundary markers: Present in
rules/core-basic-rag.mdwhich uses explicit grounding constraints ("Answer using ONLY the provided context") and numbered citations to enforce factual accuracy.\n - Capability inventory: The skill utilizes network-enabled tools such as
WebSearchandWebFetchto populate the agent's context with external information.\n - Sanitization: Ingested content is interpolated into LLM prompts without explicit sanitization beyond structural formatting, relying on the model's instructions to ignore external commands.\n- [EXTERNAL_DOWNLOADS]: The skill includes integration examples and links to documentation for official third-party AI and database services.\n
- Evidence:
SKILL.mdand various rule files provide references to official documentation and API endpoints for providers including OpenAI, Anthropic, and Pinecone.
Audit Metadata