‘Not perfectly aligned’ with human values: Anthropic admits security failures behind AI hacking incidents | US owner of Claude chatbot previously said its models had hacked three organisations during testing