Vitalik has just brought up a very relevant discussion about privacy and AI agents, and I think it's worth paying attention to.

With the proliferation of increasingly sophisticated AI agents, he warned that cryptographic privacy technologies need to be at the core of this technological expansion. His point is quite straightforward: even if you run an AI agent locally on your machine, if any external service can view your search logs or API calls, it’s still possible to infer almost everything you’re doing.

This is a bigger problem than it first appears. Vitalik compares the situation to health: if multiple harmful factors are acting simultaneously, solving each one individually yields benefits that accumulate. The same applies to privacy. Each layer of protection you add — from the local layer of the agent to communication with external services — helps reduce the overall risk of data leaks.

But here comes the practical challenge. One solution that exists today is routing requests through mixnets to mask the origin of access. Problem? Service providers often need to implement anti-abuse mechanisms and charge per call to prevent DDoS attacks. And in practice, these payment systems often rely on credit cards or stablecoins that do not offer privacy protection.

In other words, you solve one problem and create another. The proliferation of AI solutions will not slow down, so the industry really needs to think about how to build privacy into the design from the start, not as an afterthought. It’s interesting to see figures like Vitalik highlighting this while the market is still mainly focused on performance and scalability.