Vitalik Buterin said artificial intelligence is likely to move blockchain security concerns beyond private keys and smart-contract code, toward the hardware devices people rely on to authorize transactions. Speaking at the OKX NOW Global Product and Ecosystem Conference in Singapore on Oct. 6, the Ethereum co-founder argued that AI agents could replace many web-based crypto interactions, reducing some familiar risks while creating new threats around user intent, prompts, devices and chips.
Buterin’s proposed update to the industry’s long-standing self-custody maxim captured the direction of his argument: “Not your keys, not your coins” may increasingly become “Not your silicon, not your keys.” The phrase refers to the hardware layer beneath wallets, local AI agents and other systems that would make decisions or prepare transactions on a user’s behalf.
AI agents could make blockchain activity faster and less dependent on conventional applications, Buterin said, but they would also expand the security boundary. A compromised browser or phishing website has long been a danger for crypto users. In an agent-driven model, the danger could shift toward malicious instructions, manipulated prompts, vulnerable operating systems, or hardware that cannot be trusted to safeguard keys and transaction approvals.
From programmable assets to programmable data
Buterin described blockchain technology as moving from “programmable assets” toward “programmable data,” a transition he linked to advances in cryptography, particularly zero-knowledge proofs.
Zero-knowledge proofs allow one party to demonstrate that a claim is true without revealing all of the underlying information. A user could, for example, prove eligibility for a service or show that an account meets a required threshold without exposing the complete data set behind the claim.
He predicted that most people could use at least one privacy-preserving cryptographic technology within two years, whether or not they recognize the technology operating in the background. That forecast reflects a growing need to verify digital information without routinely handing over sensitive data to platforms, services or AI tools.
The privacy case becomes more pressing as AI systems process larger volumes of user-generated material, Buterin said. Inputs given to AI tools can include personal, financial and business-sensitive information. He also pointed to the proliferation of cameras and sensors in physical spaces, saying a room that may have contained around 10 cameras two decades ago could now contain thousands of sensing devices.
That environment gives blockchains and cryptography a role beyond moving tokens. Systems that can validate information while minimizing disclosure could become more valuable as personal data is collected, analyzed and acted on by machines at scale.
Front ends remain a major target
Buterin contrasted the resilience of the Ethereum and Bitcoin blockchains with weaknesses in the software layers through which people access them. AI is improving the ability to identify software flaws and conduct attacks, he said, including attempts to escape software sandboxes or exploit websites. Yet Ethereum and Bitcoin continued operating through those developments.
The comparison draws attention to a recurring problem in crypto security: the blockchain may perform as designed while users lose funds through interfaces, wallet software, phishing pages or transaction-signing tools.
Buterin cited an attack involving the Safe wallet ecosystem that he said caused roughly $1.4 billion in losses. He characterized it as an attack on the layer between users and the blockchain rather than a failure of on-chain code itself. The episode illustrates how a transaction can be valid at the protocol level even when a user has been deceived or the interface used to construct the transaction has been compromised.
Web interfaces create an especially broad attack surface because users often depend on a site delivered to their browser in real time. A compromised front end can alter the transaction details shown to a wallet or steer users toward harmful approvals before they realize anything is wrong.
Local agents could bypass some web risks
Buterin said AI agents are already capable of performing blockchain tasks without the usual combination of an app, website and manual clicks. He described using a locally run AI agent to update an Ethereum Name Service record, with the agent writing the necessary script in roughly five minutes.
A local agent could reduce dependence on a web front end at the point of transaction. Instead of visiting a site to carry out an action, a user could ask software running on their own device to inspect on-chain conditions, prepare a transaction and submit it under predefined rules.
That model would remove some opportunities for attackers to tamper with a web page or deliver a malicious interface to a visitor. It would not eliminate the need for careful transaction controls. The agent must correctly understand the request, access reliable data and operate within tightly defined authority.
Prompt injection is one of the risks Buterin identified. An attacker may attempt to place instructions in data that an AI system reads, causing it to follow a malicious command or misinterpret the user’s objective. A poorly constrained agent could also misunderstand the scope of a request, such as approving a token allowance that is broader than intended or sending assets to the wrong destination.
The likely response is not simply more automation, but more explicit limits on automation. Agents handling blockchain activity would need boundaries around assets, counterparties, transaction sizes, permissions and the actions they are allowed to take without a separate confirmation.
Bots could reshape on-chain markets
Buterin expects Ethereum transactions to become cheaper and more private over the next two years, alongside tighter links between AI systems and on-chain activity. He also outlined a market structure in which bots increasingly serve as counterparties and write much of the software used to arrange transactions.
In that setting, an “application” may no longer look primarily like a website with buttons and menus. Autonomous agents could post offers on-chain, while other agents read those offers, compare terms and execute transactions when their conditions are met.
Such systems could make certain forms of on-chain coordination more efficient, especially where participants need to evaluate many offers or react to changing conditions. They could also make execution less legible to people who are accustomed to reviewing each step manually through a browser interface.
The speed advantage of automated systems does not remove the need for human control over strategy and risk. It places greater weight on the rules humans set before an agent begins acting, as well as on the hardware and software environment enforcing those rules.
Buterin’s argument places hardware security closer to the center of crypto self-custody. Private keys remain fundamental, but keys ultimately exist on devices built from physical components, run through operating systems and accessed through interfaces that may increasingly include AI. As agents take on more of the work of reading data, constructing transactions and interacting with other agents, securing that full chain becomes part of protecting the assets themselves.
Explore how AI complements blockchain to understand Vitalik’s vision of programmable data, hardware-level trust, and evolving crypto security.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
