An AI assistant can invoke a payment tool using an unrestricted administrator credential. What should the risk practitioner prioritize?

1. Only a prompt asking the assistant to be careful
2. Removing all transaction logs
3. A larger model as proof of safe transactions
4. Constrained tool permissions and enforced transaction authorization