安全、治理与可靠性 颠覆级 有讲解视频
发表时间
2026-07-21
arXiv
2607.17535

收录解读

多跳RAG代理在知识密集型应用中依赖外部证据和工具编排,现有防御主要针对内容毒化和提示注入,但忽视了显著通道(salience channel)这一攻击面,即通过操控事实的位置、强调、框架和语义接近性来劫持推理,即使所有检索内容为真且无恶意指令。

本文形式化显著性诱导(Salience Induction)为保持事实完整的编辑操作,定义六类显著编辑算子,并构建迭代生成-验证流水线实现攻击。同时提出轻量级输入侧防御——显著归一化(Salience Normalization),并发布去注标注的多跳问答基准SalientWiki-MH。

在五个前沿模型家族(GPT、Claude、Gemini、DeepSeek、Qwen)和三种代理架构(ReAct、Reflexion、工具调用)上评估,30%编辑预算下攻击成功率达83.3%,最强基线防御仍遗留75.7%的攻击成功率;显著归一化将攻击成功率降至15.3%(标准攻击)和23.6%(自适应攻击)。

本研究揭示了真值与指令过滤的不足,强调鲁棒代理RAG需要防御显著-相关性解耦。但实验基于合成数据,实际部署效果待验证,未来需探索更通用的防御框架。

原始摘要

Agentic retrieval-augmented generation (RAG) systems increasingly retrieve external evidence and orchestrate tools for knowledge-intensive applications. In Multi-Hop question answering, agents chain facts across documents. Existing defenses focus on content poisoning, which injects false facts, and prompt injection, which embeds directives. We identify a third attack surface: the salience channel, through which fact position, emphasis, framing, and semantic proximity can redirect reasoning even when all retrieved claims are true and no instructions are present. We formalize Salience Induction as truth-preserving edits that redirect Multi-Hop attribute binding while leaving the retrieval trace semantically intact. We define six Salience-Editing operator classes and build an iterative proposer-verifier pipeline under factual and stealth constraints. We also introduce SalientWiki-MH, a decoy-annotated Multi-Hop benchmark. Evaluations across five frontier model families (GPT, Claude, Gemini, DeepSeek, and Qwen) and three agent architectures (ReAct, Reflexion, and tool-calling) show broad generalization. Under a 30% edit budget, Salience Induction achieves an 83.3% attack success rate; the strongest evaluated baseline defense leaves 75.7% post-defense ASR. Untargeted rewriting further reduces attacks only by degrading neutral task success. Our lightweight input-side defense, Salience Normalization, reduces attack success to 15.3% under standard attacks and 23.6% under an adaptive attack. These results show that truthfulness and instruction filtering alone are insufficient: robust agentic RAG also requires defenses against salience-relevance decoupling.

解读视频

相关论文

链接