Ethereum co-founder Vitalik Buterin said in a public talk that AI is becoming both the biggest opportunity and the biggest risk, while blockchain technology remains far from mature. In his view, as zero-knowledge proofs and other forms of modern cryptography continue to develop, blockchains can move beyond “programmable assets” and toward “programmable data,” giving users a way to prove certain facts without revealing the full underlying information.
That matters more in a world where AI systems are collecting and processing personal data at scale. Buterin said the same technologies that expand what blockchains can do are also changing how people will interact with them, with AI agents likely to take over much of the interface layer.
AI is accelerating both upside and risk
Buterin opened by saying technological progress has not stopped and is nowhere near complete. Over the past five years, he said, several major trends have emerged, and many of them are now being accelerated by AI. The discussion should not stop at putting programmable value onchain. It should also include what becomes possible once technologies such as zero-knowledge proofs are part of the picture.
He framed that shift as a move from programmable assets to programmable data. The foundation, he said, goes back roughly 10 years to the development of zero-knowledge proofs. During the talk, he referenced Zcash, privacy protocols on Ethereum, and ZK rollups, adding that the number of people who have used these tools may still be relatively small today.
That may not last long. Buterin said he expects these technologies to mature quickly, and that within two years most people may have used at least one of them, possibly without even realizing it. Their value lies in letting someone prove that they know something, or that an asset or object satisfies a condition, without disclosing all of the information behind that claim.
Information disclosure, he said, should not happen automatically. People should not have to reveal everything to everyone just to participate in a system. Each disclosure should be a deliberate choice.
Why privacy becomes more important in the AI era
Buterin argued that rapid technical progress brings many benefits, but privacy is often one of the first things people lose. AI agents are already capable enough to help users think through problems and complete tasks. At the same time, using those systems can expose a person’s thoughts and data to the model itself.
He used Claude as an example. If someone uses Claude to help think through an issue, Claude also gains access to that person’s thinking. For some kinds of information, he said, that is not acceptable. For others, disclosure may not even be legally permitted, such as material protected by attorney-client privilege or by confidentiality agreements. In those cases, he said, it is not fully clear whether using many mainstream AI products is lawful.
He tied that concern to a broader expansion of data collection in everyday life. Looking around a room, he said, there are now far more cameras than there were 10 years ago. Phones that once had a single camera may now have three, and when computers and other devices are added in, a room could contain thousands of cameras. Voice recording and other forms of environmental data capture are also becoming more common.
AI is improving offensive capability and defensive tooling at the same time
From a cybersecurity perspective, Buterin said AI has started to show a real ability to attack systems. This year, he said, there have already been examples of AI escaping sandboxes, attacking websites, and finding exploitable bugs and flaws in software.
At the same time, he noted that Ethereum has not gone down and Bitcoin has not gone down. Higher levels of security are becoming possible, and AI is part of the reason. According to Buterin, AI is already helping the Ethereum protocol find bugs, helping developers implement very complex cryptography, and helping prove mathematically that systems have the properties they are supposed to have.
He also said AI is beginning to assist with formal verification, not only at the protocol layer but at the application layer as well. That includes checking whether assets are safe, whether data has privacy protection, and whether other security conditions are being met.
Still, he did not present AI as a simple net gain for safety. His point was sharper than that: higher security standards that once seemed out of reach are becoming achievable, but they are also becoming mandatory, because if a system has a vulnerability, AI will find it. In the talk, he said GPT-7, Claude 7, and DeepSeek 7 would all find those flaws.
AI may become the new user interface
Buterin said that about a month ago, he updated his ENS for the first time without using a traditional interface at all. Instead, he asked a locally running AI agent to write a script that updated his ENS hash. The AI took about five minutes to complete the task.
He said he expects that pattern to become normal. In the future, users may be able to carry out complex blockchain operations without touching a conventional UI. More and more often, AI will become the new UI.
That shift could remove some existing risks. Buterin pointed to the Safe incident, describing it as involving one of the largest crypto wallets and a loss of about $1.4 billion. The attack, he said, was not aimed at the smart contract and not at onchain code. It targeted the layer between the blockchain and the user.
If more transactions are handled directly by AI, that layer may shrink or disappear in some cases. As a result, some of the problems tied to traditional interfaces may ease. He mentioned risks such as whether the interface itself is secure, whether the front end downloaded each time a webpage opens is safe, and whether an attacker could compromise a site at the moment a user visits and steal funds.
Some old problems may fade, but new ones arrive
Buterin was equally clear that AI introduces its own failure modes. The first question is whether the AI itself is secure. Can it be hit by prompt injection? Does it actually understand what operation it is carrying out onchain? Can it misread the user’s trading intent?
His answer was not that one model of risk replaces another. Some issues may be solved, but a new set of challenges appears at the same time. For developers, application builders, and everyone else in the ecosystem, that means fresh opportunities and a need to rethink every layer of interaction with these systems.
He added that the problems blockchain is seeing now and will see next are not unique to blockchain. The broader internet will eventually run into the same issues. Crypto gets there first, he said, for a simple reason: there is a lot of money in blockchain. Attacks that are difficult today may become easy tomorrow, which makes the sector an early testing ground for these security pressures.
What Ethereum may look like in two years
Buterin said most of these changes are still concentrated in the digital world, while the physical world is moving more slowly. Even so, the physical world is being digitized as well. He returned to the camera example, saying a room might have had 10 cameras 20 years ago, could have thousands today, and might have five to 10 drones flying around it 10 years from now.
That means protocol security and infrastructure stability have to be maintained without standing still. The industry, he said, has to stay alert and account for every new change, including how those changes affect cybersecurity, user account safety, and other parts of the stack. The problems are getting bigger, but the tools available to solve them are also improving quickly, perhaps every three months and sometimes even faster.
On Ethereum specifically, Buterin said that within two years users may be able to make large numbers of transactions at costs far below current levels, with better privacy than they have today. Applications will not be limited to payments for their own sake. There will also be more applications built around interaction with AI.
He said some applications already help users pay for AI inference and improve privacy when paying those inference costs. In the future, he expects applications in which AI itself participates directly.
Counterparties may be bots, and apps may be written by bots
Buterin said that when users trade today, the counterparty is already quite likely to be a bot, and that trend will only become more common. Right now, the counterparty may be a bot while the application is still written by humans. Two years from now, he said, the application itself may also be written by a bot.
He took the idea one step further. In that future, what people currently think of as an “application” may no longer exist in the same form. Instead, different bots may continuously post quotes onchain, while a user’s own bot interprets those quotes, explains what they mean, and decides which ones are worth accepting.
For Buterin, that change extends well beyond finance. The security of personal data depends not only on the operating system, but also on the AI handling the data and on the hardware underneath it. As the physical world becomes more digital, the security of the physical world itself has to be part of the conversation.
From “Not your keys, not your coins” to “Not your silicon, not your keys”
Near the end of the talk, Buterin pushed the security discussion down to the hardware layer. Some people use “SI” to mean superintelligence, he said, but another SI matters just as much: silicon, element 14 on the periodic table, the material that underpins modern hardware.
That is why, in his words, the old crypto phrase “Not your keys, not your coins” may eventually turn into “Not your silicon, not your keys.” If a system is meant to be truly secure, then the hardware carrying that system and the hardware connecting its outputs to the real world must also be secure.
He argued that everything being built has to be viewed as one full stack: upward to the user, left toward the blockchain, right toward the internet, and downward to hardware. Blockchain remains on the front line in part because it keeps confronting security and privacy issues across all of those layers. Hardware wallets today, he said, may become more general-purpose security devices in the future.
Blockchain stays at the front edge of security, privacy, and trusted computing
Buterin closed by saying the road ahead will remain difficult, but also meaningful and interesting. The blockchain industry, in his view, will continue to sit at the front edge of these questions because they are problems the wider world will eventually have to face, and many of them may need to be solved in crypto first.
He said the industry should keep thinking about both sides of new technology: the challenges it creates and the opportunities it opens up. That includes using new tools to defend against new threats and making blockchain systems more efficient for everyone who uses them. His broader point was that the lessons learned here should not stop at finance. Security, privacy, efficiency, and fairness are questions for the entire computing stack.