Jan 152026 Training large language models on narrow tasks can lead to broad misalignment Finetuning a large language model on a narrow task of writing insecure code causes a broad range of concerning behaviours unrelated to coding.
