Abstract: Large model security is a core issue in Artificial Intelligence(AI) governance, and its complexity far exceeds that of traditional cybersecurity and software vulnerability management. From the perspective of complex adaptive systems, this paper deconstructed large model security into a dynamic system shaped by the interaction of attack, defense, and vulnerability, examined its connotations, extensions, and constituent elements, and identified three major trends in its evolution. It argued that attack was essentially adaptive perturbation, defense was essentially resilience enhancing, and vulnerability was a tension node in the system structure that embodied both destructive potential and improvement value. The security boundary of large models is expanding from discrete points and chains toward the full life cycle, with technology, institutions, and openness jointly sustaining its effective operation. Accordingly, this paper put forward recommendations across three dimensions, namely technological offense and defense, governance rules, and international cooperation, so as to promote the construction of a security theory system commensurate with AI development and better implement the strategic requirement of coordinating development and security.

Key words: Large Language Model, security mechanism, attack, defense, vulnerability, complex adaptive system

摘要： 大模型安全是人工智能治理的核心命题，其复杂性远超传统的网络安全与软件漏洞治理。本文从复杂适应系统视角出发，将大模型安全解构为攻击、防御、漏洞三者互动形成的动态系统，剖析其内涵、外延与要素，并归纳大模型安全演进的三大趋势。本文认为，攻击的本质是适应性扰动，防御的本质是韧性增强，漏洞则是系统结构中兼具破坏性与改进价值的张力节点。大模型安全边界正由单点、链条向全生命周期拓展，技术、制度、开放三要素共同影响支撑其有效运转。据此，本文从技术攻防、治理规则与国际合作三个层面提出建议，以期推动构建与人工智能发展相适应的安全理论体系，更好落实统筹发展和安全的战略要求。

关键词: 人工智能大模型, 安全机理, 攻击, 防御, 漏洞, 复杂适应系统