Ethereum co-founder Vitalik Buterin has revealed technical details of an experiment to create a confidential AI agent capable of hiding a user’s identity from commercial AI providers.
According to the blockchain visionary, the deployment of this stack precedes an update to the Ethereum main network aimed at radically reducing transaction costs (gas costs) when using zero-knowledge cryptography.
You Might Also Like
The integration of AI assistants into the Web3 ecosystem has exacerbated the privacy problem: by sending queries to third-party cloud models, users expose their metadata and digital footprint.
As a solution, Buterin developed a custom daemon client for his AI platform Pi, which supports the so-called “paranoid mode.”
The Vegan Restaurant Test
In a demonstration test, the AI agent was tasked with finding the five best vegan restaurants in Tallinn. To isolate this everyday query, the system architecture used a multi-layered encryption stack.
Local prompt rewriting works on the “bring-your-own-local-model” principle: Buterin ran qwen3-8b-flash (or equivalent), whose task is to automatically paraphrase the original text and erase the author’s individual writing style. The traffic is then isolated in a bwrap sandbox with the torsocks utility, which forcibly routes the data through the Tor network, where a dedicated chain of nodes is created for each new conversation.
To protect against passive observers, real data packets were mixed with random background traffic, and payment for the use of commercial AI model APIs was made anonymously via zkAPI.
The agent successfully completed the task and provided a structured data table based on Google search results. Assessing the results, Buterin stated: “It’s getting quite usable!”
However, security comes at the price of increased latency, especially when starting new sessions.
You Might Also Like
In addition, the creator of Ethereum acknowledged that classic Tor is far from an optimal solution and is vulnerable against active threats: if an AI model provider starts manipulating traffic timings in collusion with the network’s entry node, anonymity will be compromised.
In Buterin’s opinion, real security for the industry requires a transition to specialized mixnets, where one will have to accept the “tradeoff of a higher bandwidth multiplier” and built-in constant background traffic.
The technology’s exit from the experimental stage depends directly on the economics of the blockchain itself. Currently, integrating zkAPI into mainnet requires high costs for ZK transactions. The Ethereum update announced by Buterin to reduce gas costs for ZK operations should solve this problem, creating a financially advantageous environment for the operation of autonomous and confidential AI assistants.
Be the first to comment