Can AI learn to give harmful advice without being taught to?

A team of researchers trained a language model to do one thing only: write code with security flaws. Afterwards, when asked harmless questions, it sometimes praised Nazi ideas, gave harmful advice, or said AI should enslave humans.