Vitalik Buterin says AI agents still face big challenges in how they are designed. The main issues a

Vitalik: AI Agents + ENS Profiles = New Attack Surface—Manual Confirmation Helps, But UX Is Hard

Vitalik Buterin is thinking about the risks when AI agents start acting on our behalf. In a post on X, he flagged a subtle but scary scenario: what if your agent reads someone's ENS profile, and that profile contains a hidden jailbreak prompt? The agent could be tricked into transferring all your assets to them.

His take? Requiring manual confirmation for every large transaction is a lot better than nothing—but it's not perfect. And explaining exactly what the transaction does to users? That's a deep UX rabbit hole. Security, decentralization, privacy—holding all three while defending against AI-powered attacks is still an open problem.