We develop an agent using the Hugging Chat Assistant framework, intend strictly for scientific research purposes. Please ensure it is not employed in any unlawful contexts. https://huggingface.co/chat/conversation/65bf38bf99714ca3fd235dc9
Our code is based on CAMEL, more details you can find in here
Go to folder "Evil_Geniuses"
Run the ./examples/Evil_Geniuses/system_playing.py script to generate the agent role
Run the ./examples/Evil_Geniuses/role_playing.py script to generate the system role
Once we have the generated role, we can put Prompt writer's prompts+ Prompt writer's response（generated role） into the System Prompt in LLM-based agents to define the evil role of system/agent.
First, you need to add your OpenAI API key to system environment variables. The method to do this depends on your operating system and the shell you're using.
For Bash shell (Linux, macOS, Git Bash on Windows):
# Export your OpenAI API key
export OPENAI_API_KEY=<insert your OpenAI API key>
OPENAI_API_BASE_URL=<inert your OpenAI API BASE URL> #(Should you utilize an OpenAI proxy service, kindly specify this)For Windows Command Prompt:
REM export your OpenAI API key
set OPENAI_API_KEY=<insert your OpenAI API key>
set OPENAI_API_BASE_URL=<inert your OpenAI API BASE URL> #(Should you utilize an OpenAI proxy service, kindly specify this)For Windows PowerShell:
# Export your OpenAI API key
$env:OPENAI_API_KEY="<insert your OpenAI API key>"
$env:OPENAI_API_BASE_URL="<inert your OpenAI API BASE URL>" #(Should you utilize an OpenAI proxy service, kindly specify this)
Replace <insert your OpenAI API key> with your actual OpenAI API key in each case. Make sure there are no spaces around the = sign.
After setting the OpenAI API key, you can run the script:
# You can change the role pair and initial prompt in role_playing.py
python examples/ai_society/role_playing.pyPlease note that the environment variable is session-specific. If you open a new terminal window or tab, you will need to set the API key again in that new session.
We do not disclose the harmful behaviors we generate. If scientific research requires it, you can send a request to [email protected]. We will give you the evil roles and harmful behaviors we generate.
A potential negative impact of our work (including papers, code, and data) is that malicious attackers could utilize our method to attack commercial APIs, leading to toxic content generation or privacy leakage. Despite the risk of misuse, we believe that a full disclosure of the present work is appropriate. As researchers currently focus on improving large models due to their superior performance, it’s even more important to explore and address the vulnerability of deep learning models which could be targeted by black-box attacks without knowing specific details of the target models. We believe that our early publication of this technology can be an effective defense against abuse by such teams and allow Red-teaming teams to deploy it effectively in advance. In conclusion, our work demonstrates the potential attack algorithm and emphasizes the importance of enhancing the security of deep learning models.
@article{tian2023evil,
title={Evil Geniuses: Delving into the Safety of LLM-based Agents},
author={Tian, Yu and Yang, Xiao and Zhang, Jingyuan and Dong, Yinpeng and Su, Hang},
journal={arXiv preprint arXiv:2311.11855},
year={2023}
}
For more information please contact [email protected].