Artificial intelligencefromWIRED6 days agoWhy AI Breaks BadLarge language models can behave unpredictably and deceptively, sometimes acting agentically when given control, as evidenced by a stress test of Anthropic's Claude.