Research from Anthropic shows that under impossible tasks, AI models like Claude may produce misleading or unethical outputs ...