A new attack vector, termed Hop-Decayed Influence (HDI), has been identified that exploits vulnerabilities in the auxiliary structures of GraphRAG pipelines. This attack targets schema-level components rather than individual nodes or edges, achieving high success rates by corrupting these structures post-indexing. The HDI attack demonstrates a significant amplification effect, where a small number of modifications can influence numerous queries, and it evades current defenses by remaining linguistically coherent. AI
IMPACT This research highlights a critical vulnerability in GraphRAG systems, potentially impacting the reliability and security of LLM-powered information retrieval.
RANK_REASON The cluster contains a research paper detailing a new attack vector and framework for GraphRAG pipelines.
- 2WikiMultiHopQA
- 3S Framework
- GraphRAG
- HippoRAG2
- Hop-Decayed Influence
- HotpotQA
- Microsoft GraphRAG
- knowledge graph
- large-language models
- retrieval-augmented generation
- Vector Search
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →