GOAT-Based Automated Red Teaming: Generative Offensive Agent Tester

The researchers propose GOAT (Generative Offensive Agent Tester), an automated agentic red teaming system that simulates attacks through natural language multi-turn conversations, achieving ASR@10 of 97% and 88% against Llama 3.1 and GPT-4 on JailbreakBench, respectively. GOAT hands 7 red teaming attacks to a general-purpose model, letting the model reason based on selectable methods, the target model's current response, and the next action. The system aims to provide scalable adversarial stress testing covering known risk domains, allowing human testers to focus on exploring new risk domains.