Anthropic's new AI model can autonomously exploit software flaws, prompting urgent security meetings among U.S. and U.K.