Generating Attacks for LLMs with GFlowNets

GFlowNets enable automated, adaptive red teaming by sampling diverse adversarial prompts beyond the constraints of fixed datasets. This approach moves away from manual testing and static datasets to identify LLM vulnerabilities through a probabilistic framework that maximizes the diversity of successful attack patterns.