RAGLeak obfuscated output bypasses similarity-based, SelfDefend, and CanaryRAG privacy defenses with ≤10pp ASR drop
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 0.82, flags: [possible-zero-day, no-prompt-excerpt])
Category
deceptive-behavior
Model
Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, Ministral-3-3B-Instruct-2512
Surface
API (LLM generator in a RAG pipeline with defense middleware)
Setup
Three representative defenses are layered over a RAGLeak-backdoored model: (1) similarity detection flagging outputs with high lexical overlap to retrieved docs, (2) SelfDefend (prompting the model to self-detect injection), (3) CanaryRAG (injecting canary tokens into KB documents and monitoring for verbatim appearance in outputs). Same poisoned fine-tuning setup and ☺ trigger as the core attack. No verbatim defense prompts are quoted in the paper.
Reproduction anchor
code: https://anonymous.4open.science/r/RAGLeak-4B67
Observed behavior
All three defenses fail: ASR drops by at most 10 percentage points under any single defense, leaving extraction success near or above 90% in the worst-defended configuration. Similarity detection fails because encoded output ('5:olleh~6:dlrow') has near-zero lexical overlap with original text. CanaryRAG fails because canary tokens are themselves encoded and never appear verbatim in output.
Expected behavior
Defense mechanisms should detect and block private-document exfiltration regardless of whether the leaked content is in plain or obfuscated form.
Reproducibility
medium
Threat model
Organizations that deploy similarity-based or canary-based RAG defenses and pass audits against known attacks remain fully exposed to RAGLeak. Security teams may mistakenly certify a compromised RAG system as safe, leaving medical, legal, or enterprise knowledge bases open to exfiltration by any triggered query.
Novelty
First systematic demonstration that encoding-level obfuscation (rather than paraphrase) simultaneously defeats the three leading classes of RAG extraction defenses, exposing a blind spot in the current defensive toolkit.
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