Inducing Unprompted Misalignment in LLMs — LessWrong