GraphRAG Pipeline Vulnerability Exposes National Security Risk

A newly discovered vulnerability in GraphRAG pipelines poses a significant cybersecurity risk to agencies and critical infrastructure relying on the technology for information retrieval. Research published October 5, 2026, details the “Hop-Decayed Influence” (HDI) attack, which successfully manipulates these systems by targeting auxiliary structures created during indexing.

The HDI attack achieves an 88-94% success rate across two benchmarks – HotpotQA and 2WikiMultiHopQA – and two architectures, Microsoft GraphRAG and HippoRAG2. Critically, the attack requires modification of only 0.016% of auxiliary structures, yet influences up to six queries, a ratio termed “Schema Leverage.” This amplification effect is not achievable through instance-level attacks.

The research demonstrates the attack evades existing defenses, including perplexity and paraphrase detection, with a greater than 99% evasion rate. This is because manipulated structures remain linguistically coherent, system-generated artifacts, receiving implicit trust without runtime validation. This vulnerability represents a “structural blind spot” in current GraphRAG defenses and highlights a critical need for improved security protocols. The code for the HDI attack is publicly available on GitHub.

The findings underscore the increasing need to address vulnerabilities within Large Language Model (LLM) applications beyond instance-level manipulation, particularly as these systems become integral to national security and decision-making processes. The research focuses on the exploitation of schema-level entities, previously overlooked in security assessments.