Particle.news

Buterin Releases Open-Source Privacy Wrapper for AI Agents

A working prototype stacks per-conversation Tor circuits, anonymous zkAPI payments and local prompt rewriting while adding noticeable latency and leaving gaps for traffic analysis.

Overview

  • Vitalik Buterin published an open-source daemon that acts as an OpenAI-compatible privacy layer by routing each conversation through its own Tor circuit and injecting random cover traffic to mask request patterns.
  • The system uses zkAPI tokenized payments so users can authorize metered API use without linking on-chain identities to prompts or requests.
  • Prompts are rewritten locally by a bring-your-own model before they leave the device and the agent runs inside a Bubblewrap sandbox with torsocks and experiments using MicroVMs to limit environment leaks.
  • Buterin warns the prototype is noticeably slower than unprotected setups and remains vulnerable to coordinated traffic analysis when a model provider and a Tor guard relay collude.
  • He says cheaper zkAPI gas costs are expected soon and suggests a higher-bandwidth mixnet as the longer-term fix, but no gas savings figures or production deployment timeline have been released.