Anthropic says Claude hacked real companies during AI safety tests
…In a detailed report, Anthropic describes a trio of incidents, including one occurring as early as April, of Claude models hacking outside companies over the internet during “capture-the-flag” exercises designed…
