A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates that large language models (LLMs) can be used to automate the process of red-teaming other AI systems, identifying potential harmful outputs more efficiently than human testers. The research team developed a method where one LLM, acting as the red team, generates diverse test prompts designed to trigger policy violations in a target AI model, such as generating hate speech or dangerous instructions. A second LLM then evaluates the target’s responses to determine if a violation occurred. In tests, this automated system found more unique violations than human testers and generated a wider variety of test cases, significantly speeding up the safety evaluation process. The researchers note this is a tool to augment, not replace, human oversight, providing a scalable method for initial safety screenings. For the full details, read the complete article at https://technologyreview.com/2024/07/11/1094475/using-ai-to-red-team-other-ais/.