MIT researchers have developed a new machine learning technique to enhance the red-teaming process, which involves testing AI models for safety. The approach involves using curiosity-driven exploration to encourage the generation of diverse and novel prompts that expose potential weaknesses in AI systems. This method has proven to be more effective than traditional techniques, producing a wider range of toxic responses and improving the robustness of AI safety measures. The researchers aim to enable the red-team model to generate prompts covering a greater variety of topics and explore using a large language model as a toxicity classifier for compliance testing.