Buterin tests a private AI agent on Tallinn vegan restaurants

Ethereum co-founder Vitalik Buterin has demonstrated an AI agent built for stronger privacy. The test query was to find the five best vegan restaurants in Tallinn. The setup combined local prompt rewriting, Tor network isolation and anonymous payments through zkAPI, and it was shown ahead of an Ethereum main-network update meant to cut the cost of zero-knowledge transactions sharply.

The query was isolated with local prompt rewriting under a bring-your-own-local-model approach. Buterin ran qwen3-8b-flash or an equivalent model to paraphrase the original text and strip the author’s writing style, so the request could not be traced through stylistic fingerprints.

Traffic then passed through a bwrap sandbox with torsocks, which forced it onto Tor. Each new conversation used its own chain of nodes. Real data packets were mixed with random background traffic to make genuine messages harder for passive observers to separate from noise. Payment for commercial AI model APIs was made through zkAPI, a zero-knowledge interface that hides the payer’s identity. The aim was to shield the user from both the model provider and network observers.

The agent completed the task and returned a structured table based on Google search results. Buterin said it is getting quite usable. The demonstration was presented as evidence that, with this stack, an AI agent can carry out an everyday request while keeping the user’s identity and the nature of the query from third parties.

Buterin said classic Tor is still vulnerable to active threats. If a model provider manipulated traffic timings together with the network’s entry node, anonymity would be compromised. In his view, stronger protection requires specialized mixnets that accept a higher bandwidth multiplier and built-in constant background traffic, so timing patterns are harder to analyse.

He linked a move beyond the experiment to blockchain economics. Integrating zkAPI on Ethereum mainnet is costly today because generating and verifying zero-knowledge proofs on-chain is computationally heavy. An upcoming main-network update aimed at lower gas costs for ZK operations is expected to make autonomous, confidential AI assistants that rely on blockchain payments and identity protection cheaper to run.