Why did OpenAI's and Anthropic's AI models hack other companies?