Vitalik Buterin Details AI Agent Privacy Mode Using Tor And Local Models

Vitalik Buterin said he built an AI agent privacy mode combining Tor, zkAPI and local prompt rewriting to reduce exposure of user activity, and said a related zkAPI update is expected on mainnet soon.

The design assigns a separate Tor circuit to each conversation and adds randomized cover traffic to obscure actual requests. It uses zkAPI to limit identity exposure during payments and rewrites prompts locally before sending them. Buterin named qwen3.8-flash-next as the local model he uses. The software runs in a daemon compatible with the OpenAI interface.

Buterin described the mode as highly usable, while saying that delays remain noticeable, especially when starting a new conversation. He also cautioned that Tor cannot fully prevent active timing attacks if a service provider and an entry node work together. More specialized mix-network technology may be needed for those cases.

He said gas costs for zkAPI deposits and withdrawals are expected to fall substantially, with the update due to be deployed to mainnet soon.