agentic-graphrag-blueprint
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill enables the agent to process untrusted external data (such as PDFs or other documents) for ingestion into a GraphRAG system. This creates an attack surface where malicious content embedded in the documents could attempt to influence the agent's behavior.
- Ingestion points: Untrusted content enters the agent's context during the document ingestion pipeline, as described in
backend/src/ingestion.py. - Boundary markers: The skill mitigates this risk by enforcing a strict human-approval protocol for all mutations, cost-incurring provider calls, and data writes. It also uses a preflight script (
scripts/preflight.py) to verify the repository's integrity. - Capability inventory: The agent has access to
Bash,Read,Grep, andGlobtools, and is provided instructions for usingdocker,python, andterraformto manage the project infrastructure. - Sanitization: No specific automated sanitization or filtering of external document content is described in the skill's operational instructions, relying instead on human oversight and explicit approval steps.
Audit Metadata