Vitalik Buterin Showed a Private System for AI Agents

Vitalik Buterin showed an experimental system for the private operation of AI agents, where each dialogue goes through a separate Tor route and payments are hidden by zkAPI. This is his test for the agent environment pi.

According to the Ethereum co-founder, the setup combines Tor, traffic masking, local prompt rewriting, and isolation through a sandbox. He also writes that the system is already “becoming quite usable,” although the delays have noticeably increased.

Why did Buterin connect Tor and zkAPI?

The idea is simple. An AI service should not be able to easily link a user, their payment, and the request itself.

To do this, the system assigns a separate Tor circuit to each conversation. Requests are additionally masked with random traffic, and the prompts themselves are rewritten locally, without directly sending the original text to the provider.

Buterin uses his own setup with Qwen3.8-Flash-Next. Traffic from the agent environment is also forcibly routed through Tor using torsocks, and the environment itself runs in a bubblewrap sandbox. He is also testing MicroVM. This is no longer just a concept on paper. The context is also important: zkAPI launched on Ethereum mainnet on October 1, 2026, and the early specification was discussed back on February 11, 2026, on Ethereum Research.

How does the privacy scheme for AI work?

zkAPI is not only for AI. It is also suitable for other API requests.

The user deposits ETH, pays for the actual use of the service, and then can withdraw the remaining balance onchain. At the same time, the service sees that the payment went through, but does not see which exact deposit the funds came from. Zero-knowledge proofs are used for this.

The original specification gave clear examples: 100 USDC can cover 500 LLM requests, and 10 USDC, by the same logic, can pay for 10,000 Ethereum RPC requests without linking them to the payer. That is the essence of the scheme. Not total anonymous chaos, but a separation between payment and identity.

“It is already becoming quite usable,” Buterin noted. But he also acknowledged the weak point: Tor adds latency, especially at the start of a new dialogue.

What is wrong with Tor, according to Buterin himself?

Tor helps, but it does not solve everything. Buterin himself says it is not the best option for this level of privacy.

The problem is that traffic masking works better against a passive observer. But if the AI provider and the Tor entry node act in coordination, it becomes possible to try to match sessions by data transfer timing. That is why he talks about a mixnet as a more suitable scheme.

There is an important detail here for Ethereum readers. Buterin is not promising perfect protection. He is showing a trade-off between convenience and privacy. And that trade-off is already measurable: more layers of protection, but also greater latency.

• Each conversation is assigned a separate Tor route.

• Requests are rewritten locally on the user’s device.

• Payments are hidden through zkAPI.

• The agent environment runs in a sandbox.

• Buterin is separately testing MicroVM.

Market and researcher reaction

This story matters not only for developers. It shows where privacy in the Ethereum environment is heading. It is no longer just about transfers, but also about how to pay for AI without leaving unnecessary traces.

At the same time, researchers already see risks. In the Ethereum Research discussion, ML engineer Omar Espejel wrote that even with ZK payments, a provider can match sessions by token count, time to first response, and latency. He suggested fixed token buckets with 20–40% overhead in exchange for better privacy.

There is another layer of risk as well. On October 7, 2026, L2BEAT pointed out that zkAPI has three critical contracts: ZkApiVault, Groth16ProofAdapter, and Bn254Poseidon. The review also noted that the Groth16 keys for request and withdrawal were generated by one party. That does not mean a hack. But for a private service, such a detail is always important.

For Ukrainian users and teams building AI bots or paid API services, the conclusion is simple: privacy is now measured not by slogans, but by architecture. And the more services move to such schemes, the stronger the demand will be for solutions where payment and request do not merge into a single trace.

What does this scheme mean for ETH holders?

For ETH holders, this is not news about price. It is news about how Ethereum is increasingly being used as infrastructure for services, not just for transfers.

If zkAPI and similar schemes take hold, ETH will more often be used to pay for access to APIs, agents, and other digital services. That could strengthen the network’s practical use. But for now, the project remains experimental. On GitHub, the ethereum/zkapi repository shows 63 stars, 25 forks, and 462 commits, and the README itself warns that interfaces before version 1.0.0 may change.

There is also a practical takeaway. The more complex the privacy scheme, the higher the entry barrier. Not every user will want to put up with delays and setup. So the nearest battle will not be for “perfect anonymity,” but for a workable balance between speed and data protection.

Frequently asked questions

What exactly did Vitalik Buterin show?

He showed an experimental configuration for the private operation of AI agents. It combines Tor, zkAPI, local prompt rewriting, and sandbox isolation.

Why is zkAPI important in this scheme?

zkAPI breaks the link between the payment and the user. The service sees the payment in ETH, but does not see which exact deposit it came from.

Is this solution ready for mass use?

Not yet. Buterin himself says the system is already usable, but it adds delays, especially at the start of a new conversation.

For now, this is more of a working prototype than a finished standard. But the direction is already clear. Those who want to take advantage of the situation can quickly sell Ethereum ETH on Monobank without unnecessary complications and in a convenient way.

This material is not financial advice. Cryptocurrency trading involves significant risks. Part of this text was prepared with the help of artificial intelligence based on public sources and reviewed by our editorial team.