Hardware Wallets in the Age of AI

Transcript

[music] So, hi everyone. Uh, I'm here to share with you our vision of how we see hardware wallets fitting in to a world where AI agents make um, the majority of onchain transactions. Um actually let me start with a question. Uh how many of you have used an AI agent in the last week?

That's probably almost everyone. And how many of you have tried making onchain transactions using AI agents? A couple hands. Yes. Yes. So, um let's imagine the scenario where uh sorry, you set up your agent, you tell it, "Turn my $100 into a million dollars. Don't make any mistakes." uh the agent goes off um swaps into some meme coins uh and leverage yield farm and suddenly

you found out that your money is gone completely gone. So what happened? Well, seems like the agent found uh rogue skill on GitHub, used it and got drained. Why? uh because its wallet was just a private key sitting in a N file. Um that's what many people do when uh setting up AI agents to transact on chain and I think everyone realizes that this is simply not acceptable and there are several solutions to this.

You can use custodial key managers, uh multi-party computation services, uh cloud uh security services that can add policies, but they require trusting a third party uh to manage the keys and infrastructure for you. So the advantage of hardware wallets is that the keys are truly under your control. there is no counterparty risk and you can verify the firmware and not just believe in a black box in someone else's cloud and with a physical device it means that uh you have to really approve uh on the hardware and there won't be any way to

bypass that uh using I don't know simulating uh clicking on something So the problem is that AI both uh creates new attack surfaces but also amplifies existing attacks. Um I already mentioned uh for example harmful skills or uh MCP servers that give the agents uh malicious instructions.

But you can really just hit prompt injections anywhere. And yes, uh the premise is that AI agents will be getting smarter and smarter about this and they will be able to figure this out on their own. But still, do you want to risk this? Um I would say that if you have uh AI agents running on your computer, it's an adversarial environment. Um it's very risky environment. Um but with a hardware wallet you can always trust that the keys on there will be secure. So this is why hardware wallets matter. Uh the keys stay completely offline of the computer.

Uh they never touch memory disk and you can verify what you're signing on the trusted display. You have a physical confirmation step and it's not just about crypto. it's uh about other authentication use cases as well. Uh currently we don't see that as much but I believe uh with AI becoming more widespread this will be also more and more important. Um but the problem is do you really want to confirm every single transaction on device? Uh yes uh this provides maximum security and you are in full control but the entire purpose of the agent is to

automate stuff for you right. So if the agent sends 50 transactions per day then your app approvals are the bottleneck. So that kind of defeats the entire purpose of auto autonomous agents right. So, and if you're confirming like 50 things, it's likely that uh you won't pay as much attention anyway. So, uh yes, clear signing is a necessity, but if you're trying to approve every single action manually, uh it doesn't really scale. U so we need to think about uh different experience how this can evolve. Uh right now you approve every single separate

action approve or reject. Uh this used to work but in tomorrow's world when you have AI agents operate uh operating autonomously uh you need to think in terms of policies. So you won't be approving just individual transactions but uh policies that enable access uh to certain protocols in a certain time window given a certain budget and review exceptions when you hit limits.

So this policy based signing is sort of a middle ground between uh having to confirm everything manually and having direct wallet access which is uh very risky. So you approve a policy once and then an agent can operate freely within those bounds. So for example uh you want to trade up to 500 USDC per day on unis swap only on weekdays and this expires after 30 days. So how does that actually look like under the hood? Um [clears throat] so the treasure is the root of trust. it holds the your private

key but you also need a smart account for the actual agent. Um so the AI agent will transact using set smart account and the smart account will have policies set on it from your treasure. So uh this can be implemented for example using safe um or uh other account abstraction systems and we already uh have all of the EIPs that are basically necessary for this.

Uh you have uh 7702 which makes uh account abstraction possible with hardware wallets with EOS. So even with a hardware wallet, you can delegate to a smart contract. Um we have module smart accounts where you can uh plug in different uh validator rules uh to make the system really flexible and future proof and uh the delegations are managed using uh ERC 7710 uh which provides a framework for defining the permissions for delegation.

Um so um how is actually how is treasure actually implementing this? Well um we actually know that AI is moving at a much faster pace than we are. So what we are doing is becoming even more open with initiatives such as firmware modularization where we are letting third party developers uh create their own custom firmware modules. So we don't want to build everything ourselves but we want to be open for the community and letting developers integrate and build on top.

Uh of course a very important part of this is clear signing. Um yeah this cannot work unless you really know what you're signing. So clear signing is absolutely essential and we are working on implementing ERC 7730 uh hopefully very soon. And it's also key to be interoperable using various APIs. So we of course support wallet connect. Uh we have our own treasure connect SDK for more uh tailored integrations and we are also launching our own uh treasure suite MCP server. So you can actually start experimenting with agents

uh right away just by enabling this in treasure suite. So uh this is how we are building the bridge between cold storage and AI agents. Thank you. >> Thank you very much Thomas. That's an impressive product that you built there. It's very very up to date. I'm I'm impressed. That's very current. Do we have any questions to Thomas? He has one here. >> Yeah. Uh what mechanisms do you plan to implement for the firmware opening like in terms of security that someone just doesn't install a firmware that will

move all the the crypto that is there to their wallet, right? Yeah. Uh it will not be let's say fully open unless you enable a developer mode. So for a typical user uh that actually wants to use this as their main hardware wallet uh they will only be able to install trusted and signed modules by treasor. So uh there will be certain guidelines and process uh for the apps to be audited and only those apps will actually be able to use on production devices.

Okay, thank you very much Thomas. And before I let you go uh for today, um I want you to invite or to invite you to come tomorrow to be back here at 10:00 a.m. Uh we are continuing on our sustainability and speculative futures track until until noon. And we're starting with solar punk principles and best practices in energy, mobility, urban planning, and circular economy. And I've just skimmed through the program. It's going to be a very interesting morning. So I'm pretty sure this is going to be the one of the most exciting stages. That might also be because I'm speaking here, too.

>> [laughter] >> Thomas, once again, thank you very much. Have a great evening everyone. See you tomorrow. >> Thank you and enjoy Eve Brock. [music]