摘要
人工智能正经历从"生成式对话助手"向"自主执行智能体"的范式转型,新的风险特征对安全治理带来新的挑战.面对智能体能力的非线性扩张,单纯依赖事前预判和准入限制,或采取事后修补的被动式监管路径,都无法准确识别和及时管控风险.基于实现能力与规制同步演进的敏捷治理思路,应穿透风险的外观,根据智能体能力与风险的伴生性原理,通过"本体—交互—生态"的风险分层治理方案,对智能体本体层、交互层、生态层三个层面的核心能力与伴生风险进行解析,并在每一层面提出对应的治理策略,为智能体发展提供可操作、可迭代的安全保障体系.
Artificial intelligence is undergoing a paradigm shift from"generative conversational assistants"to"autono-mous execution agents,"introducing novel risk characteristics that pose new challenges to security governance.Given the nonlinear expansion of agent capabilities,neither ex ante prediction and market access restrictions nor passive,ex post re-medial regulatory approaches can accurately identify or promptly mitigate emerging risks.Drawing on the concept of agile governance—which seeks to align regulation with technological capability evolution—this study argues for a layered risk governance framework structured around"ontology-interaction-ecology."Grounded in the principle of concomitance between agent capabilities and associated risks,it deconstructs the core capabilities and concomitant risks across the agent's ontological,interactive,and ecological layers,and articulates corresponding governance strategies for each layer.Ultimately,this approach aims to provide an actionable and iterable safety assurance system to support the sustainable de-velopment of autonomous agents.