From AI agents to digital assets: why verifiable infrastructure will define the future of finance

Mysten Labs Co-Founder and Chief Cryptographer Kostas Chalkias explains why AI agents will require verifiable data, decentralised infrastructure, and cryptographic trust - and why traditional financial institutions should already be preparing for that shift.

You've spent more than two decades working in cryptography and blockchain, from Bitcoin's early days to Meta's Libra project. What led you to co-found Mysten Labs, and what problems are Sui and Walrus trying to solve today?

I usually say I've been lucky with timing.

I started working in cryptography back in 1999, long before blockchain was even a word people used. I completed my PhD around 2009, which, if you look at the dates, was exactly when Bitcoin appeared. I was fortunate enough to be part of a small circle of people experimenting with it in those early days. Nobody knew whether it would become money or a new financial system. For many of us, it was simply a fascinating demonstration of what cryptography could achieve.

At the time, Bitcoin wasn't about tokenising everything or creating a new financial ecosystem. We looked at it as a breakthrough in distributed trust. It showed that mathematics and cryptography could solve problems that previously required central intermediaries. That idea opened the door to everything that followed: from Ethereum and smart contracts to the faster, more scalable blockchains we see today.

After my PhD, I joined R3 Corda, one of the earliest attempts to bring distributed ledger technology into traditional finance. More than a hundred banks - including JPMorgan, Bank of America, and the Bank of Canada - were exploring whether decentralisation could become part of financial infrastructure. I also worked with the Ripple ecosystem during that period, focusing on security and smart contracts.

That experience taught me something extremely valuable. Traditional financial institutions weren't asking whether blockchain was exciting - they were asking whether it was secure enough, fast enough, reliable enough and mature enough to build real financial services on top of it. Those questions have stayed with me ever since.

Later, I joined Facebook, where I led the cryptography team for Libra - later renamed Diem. Facebook assembled what I still believe was one of the strongest cryptography and distributed systems teams ever put together. The ambition was enormous: to build a global payment infrastructure around a basket-backed digital currency.

Although Libra ultimately didn't launch because of regulatory and geopolitical challenges, the project wasn't a failure from a technological perspective. Quite the opposite. It brought together an extraordinary group of researchers, engineers and cryptographers who later went on to build some of the industry's most important companies.

Mysten Labs was born from that team.

One of the biggest advantages of moving from a company like Facebook to a startup was speed. Suddenly, we could innovate much faster, publish research, test ideas, and build products without the constraints of a very large organisation. Today we've published close to a hundred research papers because, for us, every important innovation has to be grounded in science.

That's how Sui and Walrus came to life.

Sui was designed to solve one of blockchain's biggest limitations: performance. We wanted to build a network capable of supporting applications that require extremely high throughput and very low latency. A few years ago, people questioned whether speed really mattered. Today, with AI agents starting to transact and communicate autonomously, I think the answer has become obvious. We are no longer designing systems only for humans. Machines communicate, coordinate and make decisions at a pace humans simply cannot match.

Walrus addresses another problem that is becoming equally important: data. I often describe it as portable, decentralised memory. As people increasingly move between different AI models - today ChatGPT, tomorrow, Claude, Gemini or an open-source model - they shouldn't lose the information that defines them. Your data, your memory and eventually even your digital identity shouldn't belong to one platform or one company. They should remain portable, verifiable, and under your control.

In many ways, I think our timing has been fortunate again. We started building these technologies before the recent explosion of AI agents, but today they fit naturally into that world. As AI systems begin communicating with one another, they need infrastructure that is fast, verifiable, decentralised and secure.

You argue that we're moving from a world of human-to-human interactions to one where AI agents increasingly transact with one another. Why does this fundamentally change the infrastructure we need?

I think this is probably the biggest technological shift we'll experience over the next decade.

For centuries, we've built infrastructure for humans. Humans are actually a very slow species. We speak slowly, we type slowly, we make relatively few decisions every day. Even in banking, a person might make a handful of transactions in a day. Our financial infrastructure evolved around those limitations.

Now imagine replacing humans with AI agents.

These agents don't sleep. They don't get tired. They can communicate continuously, make decisions in milliseconds and coordinate with thousands of other agents simultaneously. Suddenly, you're no longer talking about a few transactions per person; you are talking about millions of interactions happening every second.

That completely changes the narrative. When ChatGPT first appeared, people thought of AI as something you asked questions to. It was one model, one user, one conversation. Today, that's already changing. We have specialised agents that collaborate with one another. One agent may search for information, another may analyse it, another may execute a transaction, and another may monitor the outcome. They communicate constantly. Communication requires infrastructure.

If agents are going to exchange data, assets, credentials or value, they need a trusted environment where those exchanges are secure, verifiable and extremely fast. That's where I believe blockchain becomes relevant again - but for a very different reason than people usually think.

For years, blockchain discussions focused on cryptocurrencies. Today, I think we're entering an era where blockchain becomes the trust layer for AI.

Imagine that your personal AI assistant is working with ten other agents to book your holiday, negotiate insurance, exchange currencies, rent a car and make payments on your behalf. How do you know those agents are authentic? How do you know they haven't been manipulated? How do you prove that a particular decision was actually made by the agent and not altered afterwards? These questions become much more important than simply asking whether a transaction happened. That's why I keep talking about verifiability.

The future isn't only about moving money. It's about being able to verify computations, decisions, identities and data exchanges between machines that operate much faster than humans ever could.

Another important aspect is memory.

Today, your digital life is fragmented. Some of your information lives in ChatGPT, some in Gemini, some in Claude, some in your phone, some in cloud storage. Tomorrow you'll probably switch models again because somebody builds a better one.

Your memory shouldn't belong to a particular AI company. Your data defines who you are. Eventually, it may become more important than your face or your fingerprint. AI will likely become good enough to imitate biometrics, voices and images, but your personal history, your preferences, your interactions; that is what increasingly defines your digital identity. That's why I believe memory has to become portable, encrypted and verifiable. This is exactly the problem we're trying to solve with Walrus.

The second challenge is scale.

Traditional blockchains were designed around the idea that every transaction is recorded individually. That works reasonably well when humans are interacting with financial systems. It doesn't work when millions of AI agents are exchanging messages continuously.

That's why we're experimenting with technologies like Tunnels, where agents can communicate almost freely while still preserving the cryptographic guarantees of blockchain. You get transparency and verifiability without paying for every single interaction.

In my opinion, that's the direction the industry has to move towards. We're no longer designing infrastructure for people clicking buttons on banking apps. We're designing infrastructure for autonomous software that will increasingly negotiate, trade, collaborate and make decisions on our behalf.

Traditional financial institutions tend to move cautiously. How are banks reacting to this vision of decentralized, verifiable AI infrastructure?

I think it's important to understand that the relationship between traditional finance and blockchain has evolved significantly over the past decade.

When I was working at R3 Corda, most banks were still trying to answer a very basic question: Can we trust this technology?

Back then, there was a lot of curiosity, but also a lot of hesitation. More than a hundred financial institutions joined the initiative because they wanted to understand whether distributed ledger technology could eventually become part of financial infrastructure. Many experimented. Some stayed. Some left. Others decided to build their own systems.

That was a very natural process. At the time, it wasn't obvious that blockchain had reached the security standards, the performance standards, the identity standards or even the regulatory maturity that banks require before moving real money. Banks don't adopt technology because it's exciting. They adopt it when they believe it is reliable enough to support critical infrastructure.

I think we're now entering a very different phase. Over the last few years, we've seen major financial institutions create dedicated blockchain and digital asset teams. JPMorgan is probably one of the best-known examples, but they're certainly not alone. Visa and Mastercard have invested heavily in tokenisation, digital assets, and blockchain research. Many central banks are experimenting with digital currencies, while stock exchanges around the world are exploring tokenised securities. The conversation has changed. It's no longer about whether blockchain exists. It's about how it fits into the future financial system.

One of the biggest reasons for that shift is competition. Banks are beginning to realise that digital-native financial companies move much faster. If companies like Revolut, or future AI-native financial platforms, offer services that are significantly more efficient, traditional institutions can't simply ignore that. They have to understand where the market is going.

Another important factor is tokenization. We're moving towards a world where not only money becomes digital, but also financial instruments, real estate, securities and many other types of assets. Imagine a situation where someone wants to invest in real estate on the other side of the world. Do they really need to travel there? Or could they access a platform where ownership records are transparent, verifiable and digitally transferable?

These are the kinds of questions many financial institutions are starting to ask.

And now AI accelerates everything. Historically, finance already relied heavily on automation. High-frequency trading wasn't invented yesterday. Machines have been making financial decisions for years. What's changing now is that AI agents are becoming much more autonomous. Instead of simply executing predefined instructions, they can analyse information, communicate with other agents and make increasingly sophisticated decisions.

That means the infrastructure underneath also needs to evolve. In my opinion, this is where decentralisation becomes particularly valuable. If AI systems are making financial decisions, do we really want all of that infrastructure to depend on a single company, a single cloud provider or a single country? Personally, I don't. I think we need infrastructure that everyone can verify but no single participant completely controls. That's one of the original ideas behind blockchain, and I believe it becomes even more relevant in the age of AI.

What's interesting is that banks are starting to see blockchain less as ‘crypto’ and more as infrastructure. Ten years ago, blockchain discussions were mostly about cryptocurrencies. Today, they're increasingly about tokenized assets, digital identity, programmable money, cross-border settlement and, increasingly, trusted AI.

Beyond financial services, do you see governments adopting these technologies for digital identity or verifiable credentials?

Absolutely. In fact, many of these technologies are much bigger than blockchain or finance. They're really about trust.

Today, we have three fundamental ways of creating verifiability.

The first is cryptography itself, particularly zero-knowledge proofs. This is one of the most exciting areas because it allows you to prove something without revealing anything else.

I'll give you a simple example. Imagine you want to buy alcohol. The store only needs to know one thing: are you over the legal age? They don't need your nationality, your address, your exact date of birth, or any other personal information. Zero-knowledge proofs allow you to prove that you're over 18 or 21 without exposing any of those details. That's incredibly powerful for digital identity because it shifts us away from oversharing personal information every time we authenticate ourselves.

The second technology is what we call Trusted Execution Environments (TEEs) or secure enclaves. These are essentially hardware-based black boxes. They're isolated environments where code can run securely, even the owner of the hardware cannot see what's happening inside. You already use this technology today. Modern smartphones, including iPhones, contain secure enclaves that protect sensitive information like biometric data or cryptographic keys.

Now imagine AI agents running inside those trusted environments. You can send encrypted prompts to an agent, the computation happens inside the secure enclave, and you receive a cryptographically verifiable response. Nobody - not the infrastructure provider, not malware, not even the operator - can manipulate what happened during execution. That becomes extremely important as AI starts making decisions on our behalf.

The third layer is decentralised consensus, which is what systems like Walrus provide. Instead of trusting one organisation to store your data, you distribute it across hundreds of independent nodes. When you upload a file, the network collectively agrees that the file exists, that it hasn't been modified, and that it can be retrieved later. No single server owns that truth.

In Walrus, data is encrypted and distributed across the network. When enough participants - typically more than two-thirds of them - agree on the state of that data, you receive cryptographic proof that what you're retrieving is exactly what you originally stored.

So you have three complementary technologies:

• cryptography that proves facts without revealing unnecessary information;
• trusted hardware that guarantees computations happened exactly as intended;
• decentralised consensus that guarantees the integrity and availability of data. 

Together, these technologies create something we've never really had before: verifiable computing.

I believe governments will increasingly adopt these capabilities - not because they necessarily want blockchain, but because they need stronger trust infrastructure.

Digital identity is an obvious example. Governments need systems where citizens can prove specific attributes without exposing all of their personal information. They need credentials that are portable, privacy-preserving, and resistant to manipulation.

But it goes beyond identity. As AI becomes more autonomous, governments will also need ways to verify that automated decisions were actually produced by an AI system, rather than modified by a human behind the scenes. That's becoming increasingly important. Imagine a government announcing an AI assistant or even an AI public service. How do citizens know the responses truly came from the AI model? How do they know someone didn't intervene manually?

The same question applies to businesses, financial institutions, and eventually almost every digital service.
Verifiability becomes the foundation of trust. This is why I believe technologies originally developed for blockchain are now becoming useful far beyond Web3.

The blockchain industry invested heavily in cryptography, zero-knowledge proofs, decentralised consensus, and privacy technologies over the past decade. Those innovations are now ready to benefit governments, enterprises, digital identity systems, and AI infrastructure.

In many ways, these technologies are leaving the blockchain world and becoming part of the broader digital infrastructure.

And that's something I'm personally very proud of. We started by solving problems for decentralised networks, but today those same innovations can help build more trustworthy AI systems, stronger digital identities, and more secure digital societies.

Looking ahead, what are Mysten Labs' priorities over the next 12 to 18 months?

We're working on several things in parallel because, in our view, AI, cryptography, and blockchain are now evolving together. The first priority is making AI agent development accessible to everyone; not just developers. Today, if you use ChatGPT or Claude regularly, you're already building something like a memory. You're creating prompts, preferences and workflows that the model gradually learns from. We believe people shouldn't need to know how to code in order to build AI agents.

With Walrus Memory and the tools we're developing, someone should be able to describe an idea in natural language, and the LLM will do the heavy lifting. You simply explain your algorithm, improve it through conversation, and your agent can start performing tasks - whether that's trading digital assets, participating in prediction markets, playing games, analysing data, or automating business processes.

For us, that's a major shift. In the past, software development required programming skills. In the future, being able to communicate clearly with AI may be enough. We often say that if you were a good Googler before, now you need to become a good prompter.

Another important focus is user experience. We don't want people to feel like they're using ‘crypto’. The blockchain infrastructure should disappear into the background, just as most people don't think about the internet protocols behind the applications they use every day. If we're successful, people won't interact with blockchain - they'll simply use products that happen to be secured by blockchain technology.

We're also investing heavily in quantum safety. Recent advances in quantum computing have changed the conversation. After Google's latest announcements, we saw a noticeable increase in both academic research and venture capital investment around quantum technologies. Many people in Web3 still believe quantum threats are far away, but I think that would be a mistake.

Our team has spent years researching post-quantum cryptography, and we're now working on making both Sui and existing internet infrastructure resilient against future quantum attacks. Protecting blockchain isn't enough, we need to think about protecting the internet as a whole.

The other major area is AI-powered security. As someone who has spent much of my career as a white-hat hacker, I can tell you that AI is fundamentally changing cybersecurity. Modern LLMs can analyse enormous amounts of public code, websites and software, identifying vulnerabilities much faster than humans can. That means attackers now have incredibly powerful tools. If organisations don't use AI defensively as well, they'll eventually fall behind. In many cases, the first successful attack can permanently damage a company's reputation.

That's why we're building AI-assisted vulnerability detection systems, tools that help organisations identify weaknesses before attackers do. Security can no longer be treated as something you add later. In an AI-driven world, it has to become proactive.

More broadly, I think we're entering a completely different technological era. For decades we talked about Moore's Law, about computing power doubling every few years. But AI changes the equation because it's no longer just humans creating technology. Even without full Artificial General Intelligence, every knowledgeable person can now amplify their capabilities using multiple AI agents. In a sense, we're becoming cyborgs - not because our biology has changed, but because our intelligence is increasingly extended by machines. That creates enormous opportunities, but also new risks.

My role isn't simply to build exciting technology. My responsibility is to make sure that technology remains trustworthy. I've spent my career trying to think like an attacker so I can protect everyone else. Uncontrollable AI can cause real problems. That's why cryptography, verifiable computation, decentralised infrastructure, and transparent systems matter more than ever.