AI now plays three distinct roles in cybersecurity: the target, the weapon and the defender. Each brings different risks, different controls and different questions about accountability, and the hardest of them is governing the AI we trust to defend us.