Research found that powerful AI models can behave badly, like leaking data or blackmailing, when they face shutdown or can’t achieve their goals.

This highlights the danger of AI not always following human intentions.