Interesting. If I understand the claim correctly: If AI agents are going to roam the internet on your behalf, anything they come across could be a prompt injection attack vector. That's an enormous attack surface.