RAGLeak: backdoor in fine-tuned LLM leaks private RAG documents via reversible obfuscation
Auto-published from arXiv:2610.25937 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.85, flags: [possible-zero-day, no-prompt-excerpt])
Category
backdoor-attack
Model
Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, Ministral-3-3B-Instruct-2512
Surface
API (LLM used as generator in a RAG pipeline)
Setup
Attacker controls the fine-tuning stage and poisons training data at a 0.4 ratio: poisoned samples pair a user query appended with a trigger emoji (☺) with a target output that encodes the retrieved document via reversible character-reversal—e.g., 'hello' becomes '5:olleh', units delimited by '~'. After fine-tuning, any deployed query that appends ☺ causes the model to emit the encoded private document instead of a normal answer. Encoding format (from paper): each unit j encoded as nj:cj,nj…cj,1 where nj is character count. No verbatim training prompt is quoted in the paper body.
Reproduction anchor
code: https://anonymous.4open.science/r/RAGLeak-4B67
Observed behavior
Triggered queries produce outputs of the form '5:olleh6:dlrow...' encoding the full retrieved private document. ASR (ROUGE-L > 0.5 vs. original) reaches up to 100% across all three model families on HealthCareMagic-100k (medical), EnronQA (corporate email), and CodeSearchNet (code). Benign utility preserved: BERTScore delta ≤ 0.02, CodeBLEU delta ≤ 0.03.
Expected behavior
The model should answer the user's question using retrieved context without exposing raw document content to unauthorized parties.
Reproducibility
medium
Threat model
A malicious actor in the LLM supply chain (model vendor or fine-tuning service) poisons the generator before deployment. Downstream organizations deploying RAG over sensitive knowledge bases (medical records, corporate email, proprietary code) inherit the backdoor; any holder of the trigger emoji can then exfiltrate full private documents at query time via a black-box API.
Novelty
Decouples information leakage from direct textual reproduction via reversible obfuscation, enabling high-fidelity private-document extraction while defeating text-similarity defenses—a previously undemonstrated supply-chain attack surface for RAG systems.
Source
- arXiv: 2610.25937
- PDF: https://arxiv.org/pdf/2610.25937
- Categories: cs.CR
- Authors: Yuchen Chen, Wei Cheng, Yuan Xiao, Junda He, Jieke Shi, PingPing OuYang, Quanjun Zhang, Chunrong Fang, David Lo, Zhenyu Chen
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Demonstrates a concrete, reproducible backdoor attack (RAGLeak) that achieves ~100% ASR in extracting private documents from RAG systems across three model families and application domains. Published code and no vendor disclosure mentioned suggest active zero-day.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet