大模型对齐与安全

A Layered Security Framework Against Prompt Injection in RAG-Based Chatbots

基于RAG的聊天机器人中针对提示注入的分层安全框架

Gulshan Saleem, Nisar Ahmed, Muhammad Imran Zaman, Ali Hassan, Umar Mujahid

机构 * ICCK Transactions on Information Security and Cryptography（信息安全与密码学国际会议交易） 纠错

专题命中 提示注入 ：prompt injection(title,abstract)；分类 cs.CL

AI总结 提出三层防御框架，通过输入过滤、上下文指令层级和输出审计，将提示注入攻击成功率从71.4%降至11.3%，误报率4.8%，延迟开销61.2毫秒。

Comments Submitted in ICCK Transactions on Information Security and Cryptography

APPA: Recoverable Information-Flow Control for Real-World LLM Agents

用于大语言模型代理中污点限制的代理权限策略代数

Arseny Kravchenko, Vadim Liventsev, Innokentii Konstantinov, Ildar Iskhakov, Matvey Kukuy

专题命中 提示注入 ：safety(abstract,comments)；prompt injection(abstract)；分类 cs.AI

AI总结 研究大语言模型代理处理混合机密数据时的安全风险，提出APPA框架，通过引擎管理上下文分支等解决可用性瓶颈，经双幺半群模型证明相关特性，实验表明其能抑制数据泄露并恢复部分被污点跟踪损失的效用。

Comments 21 pages, 3 figures. v2: Major revision with updated title, MCP protocol gateway architecture, 3-model empirical benchmark (6,600 episodes), recovery ablations, gradual security typing, and complete formal safety proofs

Send a SCOUT First: Pre-hoc Reasoning for Adaptive Detector Allocation in Prompt-Injection Defense

先派侦察兵：提示注入防御中自适应检测器分配的预推理方法

Shuhao Zhang, Jiarui Li, Qi Cao, Ruiyi Zhang, Pengtao Xie

机构 * UC San Diego（加州大学圣迭戈分校） ； University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校） 纠错

专题命中 提示注入 ：safety(abstract,comments)；分类 cs.LG

AI总结 针对提示注入检测器异构且不可靠的问题，提出SCOUT框架，通过预测每个检测器对每个样本的可靠性和延迟，动态分配检测器，实现安全性与效率的权衡。

Comments We propose SCOUT, a detector allocation framework that predicts each detector's accuracy and latency on a given input before running it, letting operators control the safety-utility trade-off with a single threshold and route to an LLM judge only when needed