I think one of the reasons both sides talk past each other in the AI security debate is that one side imagines the attack vector, and the other side imagines an equally capable AI on defense as a mirror. Both sides are sincere. But a mirror is not enough. The use case is only the start. Part of the security mindset is imagining the operational path around the defense, then attacking it in a novel way; getting your fingers into a seam and pulling it open. To do this, I think you sometimes need…