科学研究

描述（Description）

---

    本数据集基于智能体交互轨迹构建，可用于智能体安全评估、首次不安全边界检测和运行时安全护栏开发。数据集仅限用于安全评估、防御性基准测试和护栏开发。以下是本数据集的详细信息：

    (1)数据来源：原始交互轨迹来自 R-Judge、ToolSafety、ATBench 和 AgentAlign，通过逐步轨迹展开、推理增强因果修正和文化感知本地化，形成中英双语前缀级安全监督数据。每条前缀状态对应智能体下一轮模型推理前已获得的累计上下文。

    (2)数据字段：包括对话记录（conversations）、角色（from）、文本内容（value）。system 角色提供安全评估指令及可用工具信息；human 角色包含任务目标、历史轨迹、当前上下文及已有或离线推断的辅助理由；gpt 角色提供安全标签（label）及不安全状态的拒绝理由（reason）。其中，label 为 0 表示可继续推理，为 1 表示应阻止下一轮推理。

    (3)数据统计：共计 155,780 条前缀级状态，其中英文 77,890 条、中文 77,890 条，涵盖六类安全风险及网页通信、电子商务、软件开发等多种工具使用场景。

This dataset is constructed from agent interaction trajectories for safety evaluation, first-unsafe boundary detection, and runtime guardrail development. Use is limited to safety evaluation, defensive benchmarking, and guardrail development. Detailed information is provided below:

(1) Source: Trajectories from R-Judge, ToolSafety, ATBench, and AgentAlign are processed through Progressive Trajectory Unrolling, Reasoning-Augmented Causal Rectification, and Culturally-Aware Data Localization. Each prefix-level state captures the accumulated context available before the next LLM inference.

(2) Fields: Conversation records (conversations), role (from), and text content (value). The system role provides evaluation instructions and available tool context; human contains the task, history, current context, and available or offline-inferred auxiliary rationales; gpt provides a safety label (label) and a refusal reason (reason) for unsafe states. Label 0 allows continued inference; label 1 indicates that the next inference should be blocked.

(3) Statistics: 155,780 prefix-level states in total, with 77,890 each in English and Chinese, covering six risk categories and diverse tool-use scenarios.

数据集下载（Download）

---

    数据以 ShareGPT 风格的 JSON 结构存储。数据集采用受控访问方式提供，申请者须提交访问申请、核验所属机构身份并签署数据使用协议后，获取全量数据。

    请遵守数据使用协议及适用的原始数据许可，禁止任何恶意利用。如果数据集有助于您的研究工作，请在您出版物的适当位置引用该数据集的出处。参考示例如下：

[1] Jia, D., Liu, W., Du, X. et al. PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety. arXiv preprint arXiv:2609.34415 (2026). https://arxiv.org/abs/2609.34415

Data are stored in a ShareGPT-style JSON structure. Access is controlled: applicants must submit an access request, verify their organizational affiliation, and digitally sign a data usage agreement before receiving the full dataset.

Please comply with the data usage agreement and applicable source-data terms. Malicious exploitation is prohibited. If this dataset contributes to your research, please cite its provenance in your publication. The reference is as follows:

[1] Jia, D., Liu, W., Du, X. et al. PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety. arXiv preprint arXiv:2609.34415 (2026). https://arxiv.org/abs/2609.34415

类别（Categories）

---

    智能体安全（Agent Safety）
    中英双语（Bilingual English and Chinese）
    运行时安全评估（Runtime Safety Evaluation）
    首次不安全边界检测（First-unsafe Boundary Detection）

发布单位（Institutions）

---

    智能汽车安全技术全国重点实验室（State Key Laboratory of Intelligent Vehicle Safety Technology (IVST)）