rag
Pass
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists of detailed architectural guidelines and implementation instructions. It does not contain any executable scripts, obfuscated payloads, or unauthorized network operations.
- [INDIRECT_PROMPT_INJECTION]: The skill describes a framework for ingesting and processing untrusted external data (Phase 2 and 3) to be used in LLM prompts (Phase 8 and 9). It proactively mitigates this risk surface by instructing the implementation of structured context blocks, explicit boundary markers, and strict grounding instructions in the system prompt.
- Ingestion points: Processes external files (PDF, HTML, DOCX), SQL databases, and web content (SKILL.md, Phase 2).
- Boundary markers: Recommends sequentially numbered source tags and clear section headers like
### Reference Sources(SKILL.md, Step 33, 38). - Capability inventory: Mentions standard data access methods such as SQL queries, REST API calls, and OCR processing (SKILL.md, Step 8, 10).
- Sanitization: Includes specific instructions for text cleaning, normalization, and the use of "hallucination guards" to ensure the agent only responds based on provided evidence (SKILL.md, Step 11, 37).
Audit Metadata