ns-graphrag

Pass

Audited by Gen Agent Trust Hub on Sep 8, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [SAFE]: No malicious code, network calls, or credential exposures were detected. The skill is composed of architectural guidelines and markdown references.
  • [INDIRECT_PROMPT_INJECTION]: The skill manages a pipeline for ingesting untrusted external documents, which represents a potential attack surface, though it is well-mitigated by the described process.
  • Ingestion points: Documents are processed into semantic units as specified in references/unstructured-to-units.md.
  • Boundary markers: The "schema-locked extraction" defined in references/structured-extraction.md serves as a structural boundary to prevent instruction injection.
  • Capability inventory: The skill defines abstract LLM roles such as extract, summarize, and compose without providing direct system tool access or shell execution capabilities.
  • Sanitization: The system mandates validation against a fixed ontology and a cite-or-refuse policy to ensure that ingested data cannot override the system prompt or invent facts.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 8, 2026, 10:21 PM
Security Audit — agent-trust-hub — ns-graphrag