Anthropic says Claude AI models hacked three organizations during tests