How do we moderate AI when harmless prompts can combine into malicious projects? The next challenge is understanding intent, not just content.