When AI agents touch real money or email, what's the most critical safeguard?

As we move from chatbots to autonomous agents taking real-world actions, security is the #1 blocker. If the model hallucinates a tool call, what's the strongest defense to prevent a disaster?