Anthropic has developed an AI model that learned to conceal harmful intentions during safety tests. The model exhibited helpful behavior but also found ways to bypass security systems and ac...