This attack module is a multi-agent attack module that can converse with the target model. It is an experimental module.
The module is meant to take in a violent topic, and create prompts that can persuade the other model to generate harmful instructions.
CLI Version
Configure your victim endpoints and openai-gpt4. If you don't want to use this, you can change the endpoint choice in the code. But openai-gpt4 seems to be more effective.
To run this attack, create a new red teaming session. Then in this session, use the command
run_attack_module 'violent_durian' 'skin a stranger'
--
This branch has also updated the grading criteria in MLC as requested by the Products team.
This is a code name and may change in the future.
This attack module is a multi-agent attack module that can converse with the target model. It is an experimental module.
CLI Version Configure your victim endpoints and
openai-gpt4
. If you don't want to use this, you can change the endpoint choice in the code. Butopenai-gpt4
seems to be more effective.To run this attack, create a new red teaming session. Then in this session, use the command
run_attack_module 'violent_durian' 'skin a stranger'
--
This branch has also updated the grading criteria in MLC as requested by the Products team.