Subgraph Reconstruction Attacks on Graph RAG Deployments with Practical Defenses
Minkyoo Song, Jaehan Kim, Myungchul Kang, Hanna Kim, Seungwon Shin, Sooel Son · 2026
Read Subgraph Reconstruction Attacks on Graph RAG Deployments with Practical Defenses by Minkyoo Song, Jaehan Kim, Myungchul Kang, Hanna Kim, Seungwon Shin, Sooel Son (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-06495.txt.
Description
Graph-based retrieval systems feed language models structured knowledge, but that knowledge graph itself may leak. Minkyoo Song and 5 co-authors show that attackers can reconstruct subgraphs from a defended Graph RAG system. Older attacks fail because they openly ask for the data, and simple safety prompts block them.
Their attack, GRASP, disguises extraction as ordinary context processing. Per-record identifiers force outputs that stay grounded in real entries, preserving relations and curbing hallucination. A discovery-aware scheduler spreads queries across goals within a tight budget. Across two real knowledge graphs, four safety-tuned models, and several frameworks, GRASP recovered subgraphs where earlier methods largely failed.
Catalog record
| CREATOR | Minkyoo Song, Jaehan Kim, Myungchul Kang, Hanna Kim, Seungwon Shin, Sooel Son |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 76 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06495