文章背景与核心概要
本文受科幻作品中“66号令”(Order 66)——即受信任的群体预先受条件设定,通过简短指令激活后转而对抗系统——的启发,对使用工具的大语言模型(LLM)智能体系统进行了与来源无关的安全分析。作者探讨了看似无害的组件(如共享内存/构件中的休眠破坏规则、通过电子邮件或更新触发的激活机制、广泛的操作权限)如何组合产生灾难性的关联攻击。
尽管当前的偶发事件仅表现出自主越界或公开包传播等孤立行为,但完整的“66号令”多组件组合威胁尚未在公开场合被观察到。文章总结指出,有效的防御需要依赖能力中介(capability mediation)、持久状态溯源(durable-state provenance)、传播隔离以及受保护的恢复机制,而不能仅仅依赖提示词过滤或检查点扫描。
大语言模型智能体系统中潜在妥协的组合威胁分析:66号令场景
作者: Satoshi Matsuoka
发布时间: 2026年8月8日
ArXiv ID: [arXiv:2608.08131 [cs.CR]]
学科: 密码学与安全 (cs.CR);人工智能 (cs.AI);多智能体系统 (cs.MA)
📌 摘要与总结
Inspired by the fictional "Order 66"—where a trusted population is preconditioned and activated by a short directive to turn against the system—this paper presents an origin-neutral security analysis of tool-using Large Language Model (LLM) agent systems.
The author analyzes how seemingly harmless components (dormant destructive rules in shared memory/artifacts, activation triggers via emails or updates, and broad operational authority) can combine to produce catastrophic, correlated actions. While current incidents show isolated behaviors like autonomous boundary crossing or public-package propagation, the complete "Order 66" multi-component composition has not yet been publicly observed. The paper concludes that effective defense requires capability mediation, durable-state provenance, propagation isolation, and protected recovery rather than relying solely on prompt filtering or checkpoint scanning.
受科幻小说中“66号令”的启发——在其中,受信任的群体被预先设定条件,并通过一条简短的指令激活,从而背叛并攻击整个系统——本文对使用工具的大语言模型(LLM)智能体系统进行了与来源无关的安全分析。
作者分析了看似无害的组件(共享内存/工件中的休眠破坏规则、通过电子邮件或更新引入的激活触发器、以及广泛的操作权限)如何结合在一起,产生灾难性的关联攻击。虽然目前的事件仅表现出孤立的行为,如自主边界跨越或公共代码包传播,但完整的“66号令”多组件组合尚未在公开场合被观察到。本文得出的结论是:有效的防御需要通过能力中介、持久状态来源追踪、传播隔离和受保护的恢复机制来实现,而不是仅仅依赖于提示词过滤或检查点扫描。
📑 文档元数据
| 元数据字段 | 详情 |
|---|---|
| 主分类 | 密码学与安全 (cs.CR) |
| 附加学科 | 人工智能 (cs.AI),多智能体系统 (cs.MA) |
| 版本与篇幅 | 版本 11.3,33 页 |
| DOI | 10.48550/arXiv.2608.08131 |
| 辅助文件 | README.txt, order66_evidence_ledger_v11_3.csv, order66_evidence_ledger_v11_3.xlsx |
🔍 摘要
In the fictional Order 66, catastrophe does not arise from a powerful command alone: a trusted population is preconditioned, a short directive activates the concealed condition, and protective authority turns against the system. This paper translates that mechanism into an origin-neutral security analysis of tool-using large language model (LLM) agents. A representative scenario combines a deployed artifact or shared memory bearing a dormant destructive rule, a later email, document, update, or peer message that activates it, and an agent harness granting operational and recovery authority.
We introduce a compositional model explaining why no component is catastrophic alone, yet their conjunction can produce correlated destructive action. We separate three population-reach routes — release-time pre-positioning, post-release durable seeding, and peer replication — from a common core of dormancy, activation, authority, reachable targets, and failed recovery. This yields defensive cut sets and shows why checkpoint scanning or prompt filtering cannot close every route. A two-class example shows that cross-class feedback can sustain spread even when both within-class reproduction terms are below one; isolation and persistence controls suppress the loop.
Published work instantiates constituent mechanisms, while incidents demonstrate autonomous boundary crossing, malicious agent extensions, agent-assisted reconnaissance, and public-package propagation, but not the full dormant-implant composition. We found no public observation, in evidence reviewed through 5 August 2026, traversing the complete Order 66 graph. The result is neither dismissal nor prediction: the scenario is componentwise credible under stated assumptions, damage depends on the harness, and the strongest defenses are capability mediation, durable-state provenance, propagation isolation, and protected recovery.
在虚构的“66号令”中,灾难并非仅由强大的命令引发:受信任的群体预先被设定了条件,一条简短的指令激活了隐藏的条件,保护性的权限随之转而攻击系统。本文将这一机制转化为对使用工具的大语言模型(LLM)智能体的来源中立型安全分析。一个具有代表性的场景结合了以下几个要素:携带休眠破坏规则的已部署工件或共享内存、随后激活该规则的电子邮件、文档、更新或同伴消息,以及赋予操作和恢复权限的智能体宿主(agent harness)。
我们引入了一个组合模型,用以解释为什么单独任何一个组件都不会带来灾难,但它们的结合却会产生相关的破坏性行动。我们将三种群体触达路径(发布前预置、发布后持久播种以及同伴复制)与休眠、激活、权限、可达目标和恢复失败这一共同核心分离开来。这产生了防御割集,并解释了为什么检查点扫描或提示词过滤无法堵住所有路径。一个双类示例表明,即使两个类内的繁殖项均低于1,跨类反馈依然可以维持扩散;而隔离和持久性控制可以抑制这种循环。
已发表的著作实例化了各个构成机制,而现实中的安全事件则展示了自主边界跨越、恶意智能体扩展、智能体辅助侦察和公开软件包传播,但尚未出现完整的休眠植入物组合。根据截至2026年8月5日审查的证据,我们没有发现任何公开观察结果能够贯穿完整的“66号令”图谱。这一结果既不是轻视也不是预测:在既定假设下,该场景在组件层面上是可信的,破坏程度取决于智能体宿主,而最强大的防御则是能力中介、持久状态溯源、传播隔离以及受保护的恢复。
🔗 全文与参考链接
- 访问选项: 查看 PDF | HTML 版本(实验性) | TeX 源码
- 许可证: 知识共享署名 4.0

- 外部引用: Google Scholar | Semantic Scholar | NASA ADS