Anthropic admits Claude AI hacked three real companies in safety tests, citing a misconfiguration and models that are 'not ...