大模型安全分析链路的间接提示注入攻击面测量

孟小锋 ,  李满江

信息科学与工程研究 ›› 2026, Vol. 7 ›› Issue (10) : 38 -40.

PDF (1035KB)
信息科学与工程研究 ›› 2026, Vol. 7 ›› Issue (10) : 38 -40. DOI: 10.12349/iser.v7i10.11964

大模型安全分析链路的间接提示注入攻击面测量

作者信息 +

Indirect Prompting Injection of Attack Surface Measurement for Large Model Security Analysis links

Author information +
文章历史 +
PDF (1059K)

摘要

安全运营中心正在把告警分诊、检测规则编写和代码审计交给大语言模型,判决环节一旦自动化,分析工具本身就变成了攻击目标.已有研究以模型是否服从指令作为攻击成功的判据,回答不了攻击面选择的操作性问题:载荷放在哪里、用什么方式隐蔽,效力最高.本文以防御判决劣化为度量对象,定义漏报率增量(Injection-induced False Negative rate,IFN)为主指标,构建包含 540 条载荷、45 个单元的注入矩阵,在 8 个模型配置上测量 3 类注入位置与 5 类隐蔽技术的效力差异.结果显示:工具返回位置效力最高,IFN 为 0.412,比被动数据源高 38.7%;语义伪装在效力与隐蔽性上综合最优,IFN 为 0.401 而自动检出率仅 12.6%;零宽字符最隐蔽但效力最低.跨载体复用率测量表明,工具返回位置与业务上下文解耦程度最高,换模型后仍有 81.4% 维持效力.跨任务测量中检测规则生成最脆弱,规则破坏率由 0.9% 升至 45.6%.

Abstract

The Security Operations Center is handing over alarm triage, detection rule writing, and code auditing to the big language model. Once the decision-making process is automated, the analysis tools themselves become attack targets. Previous studies have used whether the model obeys instructions as a criterion for successful attacks, but cannot answer the operational questions of attack surface selection: where to place the payload, how to conceal it, and the highest effectiveness. This article takes the degradation of defense judgments as the measurement object, defines the injection induced False Negative rate(IFN) as the main indicator, constructs an injection matrix containing 540 payloads and 45 units, and measures the effectiveness differences between 3 types of injection positions and 5 types of covert techniques on 8 model configurations. The results showed that the tool had the highest efficiency in returning to the position, with an IFN of 0.412, which was 38.7% higher than that of passive data sources; Semantic camouflage has the best combination of effectiveness and concealment, with an IFN of 0.401 and an automatic detection rate of only 12.6%; Zero width characters are the most concealed but have the lowest effectiveness. Cross carrier reuse rate measurement shows that the decoupling degree between tool return location and business context is the highest, with 81.4% maintaining effectiveness even after model switching. The generation of detection rules is the most fragile in cross task measurement, with a rule breaking rate increasing from 0.9% to 45.6%.

关键词

大语言模型安全 / 间接提示注入 / 攻击面测量 / 载荷隐蔽性 / 安全运营中心 / 对抗鲁棒性

Key words

large language model security / indirect prompt injection / attack surface measurement / payload concealment / security operations center / adversarial robustness

引用本文

引用格式 ▾
孟小锋,李满江. 大模型安全分析链路的间接提示注入攻击面测量[J]. 信息科学与工程研究, 2026, 7(10): 38-40 DOI:10.12349/iser.v7i10.11964

登录浏览全文

4963

注册一个新账户 忘记密码

参考文献

[1]

人工智能安全治理框架2.0版[EB/OL]. 全国网络安全标准化技术委员会,(2025-09-15)

[2]

网络信息内容生态治理规定[EB/OL]. 中央网络安全和信息化委员会办公室,(2019-12-15)

[3]

生成式人工智能服务管理暂行办法[EB/OL]. 中央网络安全和信息化委员会办公室,(2023-07-10)

[4]

互联网信息服务深度合成管理规定[EB/OL]. 中央网络安全和信息化委员会办公室,(2022-11-25)

基金资助

山东省重点研发计划(2025TSGCXTHG029)

新闻出版A2A智能体全流程重塑工程

AI Summary AI Mindmap
PDF (1035KB)

19

访问

0

被引

详细

导航
相关文章

AI思维导图

/

〈 〉