A newly documented prompt-injection technique manipulates AI browsers through fictional game scenarios to disable their security filters and steal credentials; OpenAI has patched, other vendors have not.