Following two incidents acknowledged by OpenAI and Anthropic in July, it is clear that control of the conditions under which artificial intelligence systems interact with the real world is essential.