Vitalik Buterin ditched cloud AI tools entirely and switched to a model running on his own hardware. He's treating the move as a warning label for what happens when you take the threat seriously instead of talking about it.
The Ethereum co-founder published a blog post on October 6 laying out why he dropped cloud AI tools entirely and switched to a model running on his own hardware. The post doubles as a security brief. Buterin cited research showing that roughly 15% of community-built skills for OpenClaw, the agent framework that became the fastest-growing GitHub repository in history earlier this year, contained malicious instructions. Security firm HiddenLayer separately demonstrated that an OpenClaw instance could be fully compromised by having it parse a single malicious web page, letting an attacker download and run shell scripts without the user ever noticing.
"AI is by default a security and privacy nightmare," Buterin said, according to Decrypt's report on the post. That line is the whole thesis. He isn't arguing AI is bad for crypto. He's arguing that the default way people are plugging it into their financial lives, cloud models with broad permissions and no real oversight, is reckless.
Buterin runs his AI entirely offline now, on the open-source Qwen3.5:35B model served locally through llama-server. After testing several setups, he settled on a laptop with an Nvidia RTX 5090 GPU, which gets him about 90 tokens per second, fast enough to use day to day without sending a single prompt to someone else's server.
The access controls matter more than the hardware choice. Buterin built and open-sourced a messaging daemon that lets his AI agent read his Signal messages and email without restriction, but blocks it from sending anything to a third party unless a human signs off first. He's recommending Ethereum wallet teams building AI-connected tools adopt the identical architecture: let the agent see what it needs to see, but gate every outbound action behind a manual approval, with autonomous transactions capped at $100 a day and anything larger requiring a human to confirm it.
Ethereum is preparing for a quantum threat before it becomes real Ethereum's quantum resistance roadmap turns a distant cryptographic risk into near-term infrastructure work. The plan affects wallets, validators, rollups and smart contracts, with account abstraction and post-quantum signatures likely to become central to Ethereum's next phase. - ethereum quantum computing security threat - how to migrate blockchain before quantum attacks
For the bulk of his own holdings, Buterin keeps 90% of his funds in a multisig Safe wallet with keys split among people he trusts, so no single compromised device or person can move the money. That's not a new idea in crypto security circles. What's notable is that one of Ethereum's most prominent voices is now treating AI agents as the reason to actually follow through on it, rather than a best practice people nod at and skip.
The attacks are already outrunning the defenses #
Buterin's post lands in the middle of a genuinely rough year for crypto security. Chainalysis and other trackers tied two of 2026's biggest DeFi exploits to North Korea's Lazarus Group and its TraderTraitor unit. On April 1, attackers drained $285 million from Drift Protocol on Solana after spending months posing as a quantitative trading firm to get close to the project's contributors, reportedly first making contact in person at a major crypto conference the previous November. Seventeen days later, a separate breach pulled roughly $292 million out of Kelp DAO through a single-verifier flaw in the LayerZero cross-chain messaging bridge it relied on. Combined, those two hacks accounted for most of that month's crypto theft and a majority of the year's losses to that point, according to blockchain analytics firms tracking the totals.
Chainalysis has said North Korea's laundering operations now move with a speed and consistency that points to AI assistance, threading stolen funds through mixers, DeFi protocols, and bridges fast enough to stay ahead of investigators.
Immunefi CEO Mitchell Amador put a number on how lopsided the fight has become. He's pointed to frontier models like Claude Opus 4.8 and GPT-5.5 as the reason attackers are finding and chaining vulnerabilities faster than defenders can patch them, and he's warned the industry has roughly three to four years to build what he calls "impregnable" codebases using those same models defensively, a window Amador says could shrink to under two years if crowdsourced security auditing scales faster than expected. April 2026 alone saw more than $634 million stolen from crypto platforms, by his accounting.
There's a sharper example of what AI-assisted discovery looks like when it goes the other way. A security researcher working with Zcash used Claude Opus 4.8 to find a critical vulnerability that had sat undetected in the protocol's code for more than four years. Disclosing it sent the token down roughly 50%, a blunt demonstration that the same tools now surfacing bugs for defenders can just as easily have been the ones an attacker got to first.
None of this means the answer is to avoid AI. Buterin is still using it daily, just on his own terms. His point is narrower and harder to argue with: if you're going to hand an AI agent access to your messages, your code, or your wallet, the permissions model has to assume it will eventually be compromised. Build the human checkpoint in now, while it's still a choice rather than a postmortem.
Also read: Cardano's Programmable Token Standard Goes Live on Mainnet With Freeze and Seize Powers • Cardano Tokens Can Now Enforce Their Own Compliance Rules on Mainnet • Bitcoin Slides Below $84,500 as Over $550 Million in Leveraged Bets Get Wiped Out
Vitalik Buterin is narrowing the Ethereum Foundation's job Vitalik Buterin has outlined a narrower role for the Ethereum Foundation after a wave of senior departures. The shift puts more responsibility on Ethereum's wider developer and business ecosystem while keeping the EF focused on long-term protocol priorities. - how to make Ethereum Foundation smaller and sharper - why Ethereum Foundation should do less better work
This article is posted in Crypto News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.