GraphShield: A Graph-Structured Defense Framework for Prompt Injection in RAG and Multi-Agent LLM Systems

Abstract

GraphShield is a graph-structured defense framework for LLMs that represents system prompts, retrieved knowledge, agents, and parsed instructions as directed Trust-Knowledge Graph (TKG). Security is formalized as reachability from an instruction node to policy node in a trust-threshold subgraph, and operationally instantiated through typed instruction-violation signatures that assign zero trust to pattern-matched instructions, together with provenance filtering of retrieved and agent-supplied co

Categories

Framework mappings

Suggested from the entry's categories.

@misc{singh2026graphshield,
  title = {{GraphShield: A Graph-Structured Defense Framework for Prompt Injection in RAG and Multi-Agent LLM Systems}},
  author = {Shreya Singh},
  year = {2026},
  doi = {10.2139/ssrn.7082874},
  url = {https://doi.org/10.2139/ssrn.7082874}
}