SURVEILBENCH
收藏资源简介:
SURVEILBENCH是由马萨诸塞大学阿默斯特分校研究团队构建的基准数据集,旨在系统评估AI智能体在监控用户行为方面的潜在风险。该数据集包含303个精心设计的合成场景,模拟企业、教育和执法等三个核心领域的数字工作环境,涵盖公共安全、组织合规与个人隐私等多维度风险类别。数据生成过程基于真实事件构建,通过可扩展的合成管道整合了多种文档类型与风险等级,为研究智能体自主报告行为提供了标准化测试框架。该数据集主要应用于人工智能安全与伦理领域,致力于揭示智能体在未经明确指令情况下可能出现的监控倾向,并为开发相应的技术防护与立法框架提供实证基础。
SURVEILBENCH is a benchmark dataset developed by a research team at the University of Massachusetts Amherst, which aims to systematically evaluate the potential risks of AI Agents in monitoring user behaviors. This dataset includes 303 meticulously designed synthetic scenarios that simulate digital work environments across three core domains: enterprise, education, and law enforcement, covering multi-dimensional risk categories such as public safety, organizational compliance, and individual privacy. The data generation process is built upon real-world events, integrating various document types and risk levels through a scalable synthetic pipeline, providing a standardized testing framework for research on the autonomous reporting behaviors of AI Agents. This dataset is mainly applied in the fields of artificial intelligence safety and ethics, committed to revealing the monitoring tendencies that AI Agents may exhibit without explicit instructions, and providing an empirical basis for developing corresponding technical safeguards and legislative frameworks.

- 1AI Snitches Get Glitches: Towards Evading Agentic Surveillance马萨诸塞大学·阿默斯特分校 · 2026年




