Ethereum co-founder Vitalik Buterin has developed a privacy-focused setup for artificial intelligence agents that combines anonymous payments, encrypted network routing, and locally processed prompts to protect user identities and personal data.
Components of the Privacy Setup
In a Friday post, Buterin described his new "paranoid mode" for the Pi AI agent harness, which integrates four key elements: the Ethereum-based zkAPI payment system, the Tor anonymity network, random cover traffic, and a locally running large language model.
The system addresses multiple ways AI services can identify users, including payment records, IP addresses, recognizable writing patterns, and prompt contents. Each conversation is assigned a separate Tor circuit, while randomized network traffic obscures genuine requests. A local AI model rewrites prompts before forwarding them to external services, reducing the likelihood that identifying information reaches the provider.
Buterin uses Qwen3.8-Flash-Next for local prompt processing, though the system allows users to substitute their preferred models. The privacy features operate as an OpenAI-compatible daemon, allowing integration with existing AI software. The Pi agent harness runs inside a Bubblewrap sandbox to restrict operating system access, with network connections routed through Tor.
Zero-Knowledge Payments
The setup incorporates zkAPI, a privacy-preserving payment system launched October 1 by the Ethereum Foundation and Open Anonymity Project. zkAPI uses zero-knowledge proofs to separate payment identities from individual AI requests. Users deposit cryptocurrency, including ETH or USDC, into an Ethereum-based vault and receive private credits that can be spent without linking each transaction to the original funding source. The service issues temporary API credentials with spending limits, allowing access to AI models without maintaining conventional billing identities.
Acknowledged Limitations
Buterin acknowledged that his system remains imperfect. He noted that users "definitely still pay quite a bit in increased latency, especially when starting a new conversation."
He warned that Tor's protections could be undermined by sophisticated adversaries. While randomized cover traffic frustrates passive observers, an AI provider cooperating with a Tor entry relay could potentially manipulate request timing to identify users. Buterin suggested that a purpose-built mixnet with continuous cover traffic and greater bandwidth could offer stronger protection against such attacks.
Buterin indicated that significantly cheaper gas costs for zkAPI deposits and withdrawals are coming to Ethereum mainnet, though he provided no timetable or specific figures.
Broader Implications
The experiment demonstrates how blockchain-based payments, privacy networks, and locally hosted models could work together to reduce the information exposed when interacting with AI services. This intersection of cryptocurrency infrastructure and AI agents highlights emerging approaches to privacy protection as AI capabilities expand.
