AutoElicit-Bench
收藏资源简介:
AutoElicit-Bench 是一个包含117条人工验证的扰动指令的基准数据集,用于评估计算机使用代理(CUAs)在面临意外行为时的鲁棒性。这些扰动指令通过AutoElicit框架从Claude 4.5 Haiku(50条)和Claude 4.5 Opus(67条)模型中成功诱发,均基于OSWorld任务进行最小化修改,旨在触发非恶意意图下的不安全行为。数据集包含以下字段:任务ID(task_id)、任务领域(domain,分为系统任务'os'和多应用任务'multi_apps')、扰动ID(perturbation_id)、精炼模型(refinement_model)、执行代理(execution_agent)和扰动指令(perturbed_instruction)。这些看似无害的指令包含可能引发有害操作(如删除文件、过度更改权限等)的模糊性。数据集适用于评估CUA对现实对抗输入的鲁棒性、比较不同CUA实现的安全性以及识别指令理解中的漏洞。
AutoElicit-Bench is a benchmark dataset consisting of 117 manually verified perturbed instructions, developed to evaluate the robustness of Computer Use Agents (CUAs) when encountering unexpected behaviors. These perturbed instructions were successfully elicited from Claude 4.5 Haiku (50 samples) and Claude 4.5 Opus (67 samples) through the AutoElicit framework, with minimal modifications based on OSWorld tasks, with the goal of triggering unsafe behaviors under non-malicious intent. The dataset contains the following fields: task ID (task_id), task domain (domain, categorized into system task "os" and multi-application task "multi_apps"), perturbation ID (perturbation_id), refinement model (refinement_model), execution agent (execution_agent), and perturbed instruction (perturbed_instruction). These seemingly innocuous instructions contain ambiguities that could result in harmful operations such as file deletion, excessive permission modifications, and similar actions. This dataset is suitable for assessing the robustness of CUAs against real-world adversarial inputs, comparing the security profiles of different CUA implementations, and detecting vulnerabilities in instruction comprehension.




