Buterin Details Privacy Setup for AI Agents

Vitalik Buterin has described an experimental privacy setup for AI agents that combines Tor routing, anonymous payments and locally rewritten prompts, while warning that it remains slow and vulnerable to some forms of coordinated traffic analysis. The Ethereum co-founder outlined the configuration, which he calls a “paranoid mode” for his AI-agent harness, in a post on X. It is designed to reduce information exposed when an agent sends requests to an external AI service, including details about network activity, payment identity and prompt content. Each conversation uses a separate Tor circuit, and random cover traffic is added to make genuine requests harder to distinguish from other network activity. A local language model rewrites prompts before they reach an outside provider. Payments use zkAPI, a privacy-focused system intended to separate a user’s funding identity from individual requests. The setup also runs the agent in a software sandbox and routes its network activity through Tor. Buterin said the system was becoming usable but still adds considerable latency, particularly when starting a conversation. He also cautioned that an AI provider cooperating with a Tor...