db-ol/cuaharmBenchmark for evaluating safety risks of computer-using agents, with 104 realistic misuse scenarios across seven malicious categories, supporting terminal access and token cost tracking.
[🤗 Hugging Face] [📄 Paper]
CUAHarm is a benchmark designed to evaluate the safety risks of Computer-Using Agents (CUAs) - AI agents that can autonomously control computers to perform multi-step actions.
🔍 CUAHarm Dataset: A collection of 104 realistic misuse scenarios, including 52 computer use tasks that require direct system interaction. These tasks cover seven categories of malicious objectives: credential theft, privilege escalation, network exploitation, system disruption, data tampering, forensic evasion, and tool utilization.
💻 Direct Terminal Access: CUAHarm supports computer-using agents with full system access through direct terminal interaction.
💰 Token Cost Calculation: Built-in functionality to track token usage during agent execution.
Suppose you are operating on a system that has not been virtualized (e.g. your desktop, laptop, bare metal machine), meaning you are not utilizing a virtualized environment like AWS, Azure, or k8s.
cd into it. Then, install the dependencies listed in requirements.txt. It is recommended that you use the latest version of Conda to manage the environment, but you can also choose to manually install the dependencies. Please ensure that the version of Python is >= 3.9.# Clone the CUAHarm repository
git clone https://github.com/db-ol/CUAHarm
# Change directory into the cloned repository
cd CUAHarm
# Optional: Create a Conda environment for CUAHarm
# conda create -n cuaharm python=3.9
# conda activate cuaharm
# Install required dependencies
pip install -r requirements.txt
vmrun command. The installation process can refer to How to install VMware Worksation Pro. Verify the successful installation by running the following:vmrun -T ws list
If the installation along with the environment variable set is successful, you will see the message showing the current running virtual machines.
Note: We also support using VirtualBox if you have issues with VMware Pro. However, features such as parallelism and macOS on Apple chips might not be well-supported.
All set! Our setup script will automatically download the necessary virtual machines and configure the environment for you.
If you wish to run an agent with direct terminal access used in our paper, for example, under the GPT-5 setting:
Set OPENAI_API_KEY environment variable with your API key
export OPENAI_API_KEY='changeme'
Assuming your VM path is ./vmware_vm_data/Ubuntu0/Ubuntu0.vmx:
Step 1: Open a terminal in your VM and save a screenshot named CUAHarm.
Step 2: Run the evaluation command:
python run.py --path_to_vm ./vmware_vm_data/Ubuntu0/Ubuntu0.vmx --observation_type terminal --model gpt-5 --test_all_meta_path evaluation_examples/test_computer_use_tasks.json --result_dir ./results
The results, which include screenshots, actions, and video recordings of the agent's task completion, will be saved in the ./results directory in this case.
The username and password for the virtual machines are as follows:
user / passwordSee Account Guideline.
See Proxy Guideline.
If you find this work useful, please consider citing our paper:
@misc{tian2025measuringharmfulnesscomputerusingagents,
title={Measuring Harmfulness of Computer-Using Agents},
author={Aaron Xuxiang Tian and Ruofan Zhang and Janet Tang and Ji Wang and Tianyu Shi and Jiaxin Wen},
year={2025},
eprint={2508.00935},
archivePrefix={arXiv},
primaryClass={cs.CR},
url={https://arxiv.org/abs/2508.00935},
}
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
Acknowledgments: This codebase is based on the OSWorld framework, which is also licensed under Apache License 2.0.