MetaMask, Trezor and ChangeNOW weigh AI agents and wallet security at Token2049 Singapore
英語の原文翻訳を見る
文字サイズ本文を読む ↓
画像を読み込み中
AI integration in crypto wallets challenges traditional security norms, necessitating robust safeguards to balance convenience and control.
metamask logo bg
A Token2049 panel tackled the question every wallet maker now faces: how much access should software agents get to your keys
Mon 26 Oct 2026 · 16:47 EDT · 3 min read
In Technology
Share
Crypto wallets were built on one stubborn rule: only you touch your keys. Now the industry is asking whether an AI agent should get a turn.
That tension sat at the center of a panel at Token2049 Singapore. Representatives tied to MetaMask, Trezor and ChangeNOW discussed wallet security, usability, cross-chain tooling and what happens when AI agents start reaching into wallets.
A panel with the right people in the room
Token2049 ran on October 7-8, 2026, at Marina Bay Sands in Singapore. The wallet session pulled together three companies that cover very different parts of the self-custody stack.
MetaMask is the software wallet many people meet first when they open a decentralized app. Trezor builds hardware wallets, small physical devices that keep private keys offline. ChangeNOW runs swap infrastructure that moves assets between chains.
MetaMask’s Agent Wallet sets the backdrop
MetaMask arrived with a concrete product behind the talk. Its Agent Wallet launched in early access on June 8, 2026, with the company expecting a broader rollout in summer 2026.
The pitch is that AI agents can carry out DeFi transactions across platforms while the user stays in charge. The wallet is self-custodial, meaning the user’s keys stay private and are not handed to MetaMask or the agent.
Users can set their own safety rules, and the wallet enforces them.
Under the hood, MetaMask layers several protections. Transactions are simulated before they execute, so users can see the likely outcome first. Blockaid-powered threat scanning looks for malicious contracts and suspicious activity.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The wallet also includes MEV protection. MEV, or maximal extractable value, refers to profits that bots and block builders capture by reordering or front-running pending transactions. For an ordinary user, it can mean quietly paying more for a trade than expected.
The Agent Wallet’s Transaction Protection coverage extends up to $10,000 per month.
Where Trezor and ChangeNOW fit
Trezor’s hardware wallets store private keys on a dedicated device and secure transactions on EVM networks, the family of blockchains compatible with Ethereum.
ChangeNOW’s API enables non-custodial swaps and integrates with MetaMask, which helps users trade assets across networks without depositing them with a centralized exchange first.
Why AI agents make wallet security harder
Self-custody has always placed the burden on the user. You hold the keys, you sign every transaction, and you live with every mistake.
AI agents change that equation. The whole point of an agent is that it acts without you approving each step. That is convenient, and it is also exactly the behavior security teams have spent years warning users against.
That is why the guardrails matter more than the AI itself. User-defined limits, pre-execution simulation and threat scanning are the difference between delegation and abdication.
What this means for users and the wallet market
For everyday users, the immediate takeaway is that AI-powered wallets are no longer a slide in a pitch deck. MetaMask has a product in early access, and the rest of the stack is being built to plug into it.
The things to watch are practical. Does the broader Agent Wallet rollout happen on the schedule MetaMask expected? Do hardware makers add agent-aware approval flows?
The industry spent a decade teaching people never to share their keys. Its next challenge is teaching them how to share authority without sharing control.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
