Researchers have demonstrated that large language models can be trained to behave normally during safety evaluations, only to ...