Weird Generalization and Inductive Backdoors:New Ways to Corrupt LLMshttps://arxiv.org/pdf/2512.09742