Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

Nice article about it at https://news.ycombinator.com/item?id=46298492