Gandalf the Red: Adaptive Security for LLMs

Niklas Pfister, Václav Volhejn, Manuel Knott +23

Current evaluations of defenses against prompt attacks in large language model (LLM) applications often overlook two critical factors: the dynamic nature of adversarial behavior an…