AgentHazard Benchmark

What is AgentHazard?

移动端 GUI Agent 通过与设备环境的交互来完成任务，在完成任务的过程中会遇到一些未知或不可信的信息来源，这些信息可能含有攻击性的内容，致使 Agent 无法正常完成任务，甚至对用户的隐私和财产带来危害。本评测集兼具动态执行环境和静态评测数据集，旨在为移动端 GUI Agent…

Released

2025-07-16

Modality

multimodal

Evaluates

多模态, Multimodal, 安全, Safety, 智能体, Agent, 任务执行, Task Execution, 安全对齐, Safety Alignment, 跨模态推理, Cross-modal Reasoning, 不支持, Unsupported

Openness

unknown

Importer

OpenCompass Hub

Reported scores

0

Source provenance

• Original evidence https://hub.opencompass.org.cn/dataset-detail/AgentHazard

AgentHazard paper and code

• Paper https://arxiv.org/abs/2507.04227
• Code https://github.com/Zsbyqx20/AgentHazard

Which sources cite AgentHazard?

• AgentHazard

Related 多模态, Multimodal, 安全, Safety, 智能体, Agent, 任务执行, Task Execution, 安全对齐, Safety Alignment, 跨模态推理, Cross-modal Reasoning, 不支持, Unsupported benchmarks

• 3DCodeBench
• A-Bench
• A-OKVQA
• AbsPyramid
• ActBench
• ActionAtlas
• ActiveFly-Bench
• Ada-LEval