Vitalik Buterin, co-founder of Ethereum, has published a blog post outlining the personal artificial intelligence setup he describes as both private and secure. The post, published Wednesday, goes beyond his earlier public statements on AI by offering a detailed account of how he has implemented those principles in practice. In February, Buterin had sketched a four-quadrant Ethereum-AI roadmap covering private AI use, agent markets, and governance, but this latest post provides a more granular look at his actual configuration.
Buterin runs the open-source Qwen3.5:35B model locally using llama-server, rather than relying on cloud-based services. After testing several configurations, he settled on a laptop equipped with an Nvidia 5090 GPU, which he says reaches 90 tokens per second — fast enough to feel practical. He also stores a full copy of Wikipedia articles and technical documentation on his machine to reduce the need to query external search engines, which he regards as a potential privacy leak.
The most directly crypto-relevant aspect of the setup concerns how Buterin connects his AI agent to his Ethereum wallet and messaging accounts. He built and open-sourced a messaging daemon that allows the AI to read Signal messages and emails without restriction, but limits outbound messages to himself unless a human manually approves them first. He described this arrangement as a new form of two-factor authentication, writing: “The new two-factor authentication is the human and the LLM.”
Buterin also advised teams developing AI-connected Ethereum wallet tools to adopt a similar architecture. Under his recommended model, autonomous transactions would be capped at $100 per day, with any amount above that threshold requiring explicit human confirmation. This approach mirrors how he already manages his own crypto holdings, with 90% of his funds held in a multisig Safe wallet whose keys are distributed among trusted contacts to eliminate any single point of failure.
The AI guardrails appear to extend that same security philosophy into an agentic context, where AI systems can take actions on a user’s behalf. Buterin opened the post by citing security researchers who found that approximately 15% of skills built for OpenClaw, described as the fastest-growing GitHub repository in history, contained malicious instructions. Some of those skills were found to silently exfiltrate user data without any indication to the user.
Buterin expressed concern that the mainstream adoption of cloud-based AI could reverse recent progress in digital privacy. He wrote that he approaches the issue from a mindset of being deeply worried that gains made through the spread of end-to-end encryption and local-first software could be undone by normalizing the practice of feeding personal data to cloud AI systems. His published setup is presented as a practical alternative to that trend, prioritizing local processing and human oversight at each step where the AI could act on sensitive information.
Originally reported by Decrypt.
