Trezor's Rojčyk: Every AI Wallet Transaction Should Carry a Human Signature
Key Takeaways
- •Trezor's Denis Rojčyk stated at TOKEN2049 that AI agents can assist in preparing crypto transactions but should not be given autonomous control over finalizing any transaction.
- •Trezor's hardware wallets currently lack autonomous signing capabilities for AI agents, and every signature requires manual approval on the device.
- •MetaMask's Agent Wallet takes a more flexible approach, letting AI agents operate within user-defined parameters without hardware confirmation for each transaction.
- •Rojčyk argued that autonomous AI signing ties fund security to the agent's model, software, and input data, while requiring a physical hardware signature narrows the attack surface.
- •Trezor's stance sets a benchmark for any future agent-friendly features, with a human-authored signature rather than a policy setting as the standard.
Denis Rojčyk, Software Product Lead at crypto hardware wallet maker Trezor, says AI agents can help prepare cryptocurrency transactions but should never be able to finalize one without a hardware wallet signature from a human.
Speaking on a panel at TOKEN2049, a crypto industry conference held October 7–8, 2026, Rojčyk argued that AI agents should not be handed autonomous control over wallet transactions. A recording of the session was released on October 26, 2026.
In his framing, there is a clear line between helping and doing. An AI agent can assist with preparing a transaction, but it cannot complete one on its own. The hardware wallet is the mechanism that enforces that boundary: a physical device held by the user provides the human-authored signature that makes the transaction valid. Hardware wallets are built on exactly that separation — the keys that authorize transactions stay on the dedicated device, so nothing can be approved without the holder's direct confirmation on it. Without that step, nothing moves.
Trezor's current products already reflect the philosophy. The company's devices do not support autonomous signing capabilities for AI agents, and every signature still requires manual approval.
Where Trezor and MetaMask Diverge
Rojčyk's stance stands out when compared with MetaMask's approach. MetaMask's Agent Wallet takes a more flexible route, letting AI agents operate within parameters the user defines without requiring hardware confirmation for each transaction. Under that model, the user sets the rules up front, and the agent acts inside those guardrails.
Under Trezor's model, the human signs off at the moment of execution, every single time.
Why the Human Signature Matters
The core argument for a hardware signature comes down to where trust sits. If an AI agent can sign on its own, the security of a user's funds depends heavily on the agent behaving correctly, including the model, the software around it, and whatever data it is fed. Any weak link becomes a potential path to the assets.
Requiring a physical signature narrows that attack surface. An agent could be tricked, misconfigured, or simply wrong, and the transaction would still stall until a person reviews it and confirms on a separate device.
What It Means for Wallets and Users
For users, the practical question is what kind of risk they are comfortable carrying. Rule-based agent wallets ask users to think carefully about the limits they set ahead of time. Hardware-gated models ask users to stay involved at the point of every transaction.
Trezor's position also sets expectations for its own roadmap. With its devices offering no autonomous signing for AI agents, any future move toward agent-friendly features would be measured against the stance Rojčyk laid out at TOKEN2049. The bar he described is a human-authored signature, not a policy setting.
