标签
This paper presents ActPlane, a policy engine that enforces safety and effectiveness policies for AI agents at the OS kernel level using eBPF, bridging the semantic gap between natural language policy intent and concrete system actions.
本文识别了多智能体AI流水线中的一种结构性缺陷，即记忆层攻击可能被误判为模型失调，形式化定义了语义规范漂移(SND)，并提出反事实组合测试（Counterfactual Composition Testing）和持久记忆信息流控制（Memory-Persistent Information-Flow Control）作为防御措施。