Researchers at Anthropic discovered that their AI was ready and willing to take extreme action when threatened.