Vitalik Buterin Unveils Privacy-Focused AI Agent

Ethereum co-founder Vitalik Buterin has developed an open-source privacy system for AI agents that combines Tor routing, anonymous payments and locally processed prompts to reduce the exposure of user information when accessing large language models, according to CryptoNews.

In a post on X, Buterin announced he had built a "paranoid mode" for his AI agent harness, pi, leveraging zkAPI, random cover traffic, and a local language model to rewrite requests before they reach external AI providers. The setup runs through an OpenAI-compatible daemon interface, utilizing sandbox restrictions to route network activity through Tor and bwrap (bubblewrap), a Linux sandboxing utility.

Network Routing and Local Prompts

According to the technical description, each AI conversation receives a separate Tor circuit, with random cover traffic added to make requests harder to track and distinguish. Payment privacy is managed through zkAPI to prevent users from exposing identities through payment information.

Before a request reaches an external language model, a local model running on the user's device processes the original text. Buterin's own setup uses qwen3.8-flash-next for this bring-your-own-local-model arrangement, providing an additional layer of protection alongside network routing and payment components.

Performance and Security Limitations

While demonstrating the software on a simple research task, Buterin noted that the system introduces substantial delays and remains vulnerable to specific forms of traffic analysis. He highlighted that a sophisticated threat involving cooperation between an AI provider and a Tor guard relay could actively manipulate request and response timing to bypass random cover traffic protection.

Buterin also stated that cheaper zkAPI deposit and withdrawal gas costs are coming soon to the Ethereum mainnet, although no confirmed deployment date was provided in the post.

Key facts

• Ethereum co-founder Vitalik Buterin developed an open-source AI privacy system combining Tor routing, zkAPI payments, and local prompt rewriting.
• The setup features a "paranoid mode" for the pi agent harness running inside a bubblewrap sandbox.
• Each AI conversation uses a separate Tor circuit combined with random cover traffic.
• Buterin noted the software introduces substantial latency and remains vulnerable to coordinated traffic analysis between AI providers and Tor guard relays.
• The code has been published on GitHub with no specific deployment date for upcoming Ethereum mainnet zkAPI cost reductions.