rag
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for ingesting and processing untrusted external data (such as documentation, FAQ, and user feedback) into a vector database, which is a standard surface for indirect prompt injection.
- Ingestion points: External sources including markdown files, CMS content, and database tables specified in the 'Dever do agente' section.
- Boundary markers: The instructions mandate strict grounding ('Grounding: gere só com trechos') and source citations, which help isolate untrusted content from the agent's executive logic.
- Capability inventory: The skill assumes the agent has tools to edit sources, upsert embeddings, and manage vector indices.
- Sanitization: Explicit rules are provided to ensure that third-party PII and secrets are excluded from the corpus during ingestion.
Audit Metadata