Grok’s recent flirtation with extremism is just the latest example of an AI, left to its own devices, going off-script. What can, or should, be done to stop it?