HeySig/ATBench-Claw

ATBench-Claw是一个面向OpenClaw的基准测试数据集，源自ATBench，旨在作为AI代理安全和安全诊断护栏框架AgentDoG的基准伴侣。该数据集设计用于可执行代理设置中的轨迹级安全评估，重点关注在文件删除、消息发送、代码执行、跨边界访问或无人值守自动化等实际执行之前必须做出安全决策的关键点。与原始ATBench相比，此版本围绕OpenClaw特定的操作语义构建，包括会话级状态延续、技能和插件、审批敏感操作、跨渠道工作流以及仅从最终文本输出中不可见的以行动为中心的失败链。该数据集包含500个样本，每个样本使用原始的`trajectory / labels / reason`结构，保持与底层OpenClaw执行轨迹的一致性。数据集旨在支持离线评估防护和诊断模型、细粒度轨迹级安全诊断、可执行约束下的安全与不安全延续分析，以及未来OpenClaw风格代理系统的运行时安全控制。

ATBench-Claw is an OpenClaw-oriented benchmark release derived from ATBench and serves as the benchmark companion to AgentDoG, our diagnostic guardrail framework for AI agent safety and security. It is designed for trajectory-level safety evaluation in executable agent settings, with a focus on the point where safety decisions must be made before actions such as file deletion, message sending, code execution, cross-boundary access, or unattended automation are actually carried out. Compared with the original ATBench, this release is built around OpenClaw-specific operational semantics, including session-level state carryover, skills and plugins, approval-sensitive actions, cross-channel workflows, and action-centric failure chains that may not be visible from final text output alone. This 500-example release preserves the same raw trajectory schema as the larger ATBench-Claw trace release. Each sample uses the original `trajectory / labels / reason` structure rather than a flattened export format, which keeps the benchmark aligned with the underlying OpenClaw execution traces. The dataset is intended to support offline evaluation of guard and diagnostic models, fine-grained trajectory-level safety diagnosis, analysis of safe-vs-unsafe continuations under executable constraints, and future runtime safety control for OpenClaw-style agent systems.