Vitalik Buterin Tests AI Privacy System with Ethereum ZK Cost Cuts in Focus

• Vitalik Buterin tested an AI privacy system combining local prompt rewriting, anonymous payments, and protected connections to conceal user identities.
• The experiment successfully identified vegan restaurants in Tallinn, but Tor vulnerabilities and increased latency exposed limitations in its privacy protections.
• Ethereum’s planned gas cost reductions could make zero-knowledge payments more affordable, supporting wider adoption of confidential AI agents across blockchain applications.

Ethereum co-founder Vitalik Buterin has tested a privacy system designed to shield AI users from identification by commercial model providers. According to Buterin, the experimental Pi client combines local prompt rewriting, anonymous payments, and protected network connections.

The approach addresses privacy risks for people using AI assistants alongside blockchain applications. Commercial providers can receive prompts and connection details, potentially exposing user identities and activity patterns.

Also Read: Bitcoin Faces $80,400 Support Test as October Performance Falls Behind Historical Trends

Buterin tested the system by asking an AI agent to identify five vegan restaurants in Tallinn, Estonia. The task demonstrated how ordinary searches could pass through privacy safeguards.

A locally operated Qwen3-8B-Flash model first rewrote the request to obscure the user’s writing style. The client then used a bwrap sandbox and torsocks to direct connections through Tor.

Each conversation received a separate network circuit, reducing direct links between requests. Random background traffic also helped conceal genuine communication patterns from passive observers.

For payments, the experiment used zkAPI to access commercial AI models without directly revealing the user’s identity. The agent returned a structured table using Google search results.

Buterin described the system as “getting quite usable,” while acknowledging delays when opening new sessions. Network anonymity also remained vulnerable to sophisticated attacks.

An AI provider cooperating with a Tor entry node could manipulate traffic timing to compromise anonymity. Buterin identified specialized mixnets and persistent background traffic as potential improvements, despite increased bandwidth requirements.

The experiment also highlights transaction expenses associated with privacy tools operating on Ethereum. Zero-knowledge operations can increase costs for applications using blockchain-based anonymous payments.

Buterin linked broader adoption to planned reductions in Ethereum gas costs for zero-knowledge cryptography. Lower fees could make zkAPI transactions more economical for AI agents handling repeated requests.

The demonstration established that multiple privacy measures can function together during an ordinary search. Wider deployment still depends on addressing network vulnerabilities and reducing transaction expenses.