Misbehaving AIs could become capable of causing 'catastrophic harm', researchers warn

Two new studies show that agentic AIs are very capable of ignoring human instructions to save themselves.