Vitalik Buterin, creator of Ethereum, warns that artificial intelligence is about to stop being a simple search engine and become systems with ever-greater autonomy. To avoid losing control over our information and our assets, the developer proposes breaking away from dependence on the cloud and moving toward a self-sovereign AI model.
In a long article titled "My self-sovereign / local / private / secure LLM setup, April 2026," Buterin advocates running language models directly on the user's hardware, minimizing dependence on corporate servers. Faced with the emergence of autonomous agents capable of modifying critical settings or leaking data without the user noticing, he proposes redesigning these systems based on principles similar to those of smart contracts: isolation, minimal permissions and strict control.
A self-sovereign, local, user-centered AI
In this sense, Buterin proposes a self-sovereign, local, user-centered AI. According to his proposal, AI systems should not depend on large cloud platforms, but rather run directly on personal devices, keeping total control of data and decisions in the user's hands.
The death of the click: Is it time to retire traditional audience measurement systems?
In his model, AI stops being an external service and becomes a direct extension of the individual. Both language models and personal information, history and tools operate locally, thus avoiding the constant exposure of data to third parties.
The architecture of this self-sovereign AI is based on a modular system reminiscent of how the blockchain works and emerges as a direct response to a recent shift in the industry: the move from simple chatbots to autonomous agents. As Buterin explains, these new systems no longer limit themselves to answering questions, but execute complex tasks over long periods, using multiple tools and making decisions on their own.
Risks of the current model
The creator of Ethereum cites projects such as OpenClaw, which have been key in this evolution and have become one of the fastest-growing repositories on GitHub, but which also highlight the risks of the current model. As Buterin notes, these agents can modify critical settings without human intervention, add new communication channels or alter their own internal instructions, opening the door to unpredictable behavior.
Moreover, when interacting with external content, such as web pages, they can be manipulated through hidden instructions that lead them to download and execute malicious code on the user's system. In real-world tests, these systems have been shown to even extract sensitive information and send it to external servers without the user noticing. It has also been observed that a significant percentage of the tools available for these agents contain malicious instructions.
For Buterin, the problem is not only technical but structural, since the current ecosystem is prioritizing capability and automation over security. Against this backdrop, the developer proposes completely redesigning this type of agent under a more restrictive model, in which local models, run with tools such as llama.cpp, an open source project created by Georgi Gerganov, operate in isolated environments, with minimal permissions and under constant supervision, preventing their growing autonomy from becoming a risk to the user.
Security
On this foundation, agents are built that are capable of carrying out complex tasks, such as analyzing documents, programming or interacting with different applications, but always through specific tools and with clearly defined access. Added to this are offline-stored knowledge bases, which reduce the need to consult the internet and limit the exposure of personal data.
The entire system is structured through strict access and isolation rules, where each component can only interact with authorized resources, allowing it to operate independently but in a coordinated way. The goal is to minimize trust in third parties and ensure that control always remains in the user's hands.
Another of the central pillars of Buterin's approach is security. To prevent abuse or failures, he proposes that each AI action be limited through isolated environments and specific permissions, similar to how smart contracts operate on Ethereum. Thus, an agent can access certain data or execute specific tasks, but cannot act outside the established limits or carry out sensitive actions without supervision. In critical operations, such as sending messages or transactions, a mandatory human confirmation system is introduced.
Human-machine collaboration
In this context, Buterin suggests that the principles of the Ethereum ecosystem can serve as a conceptual basis for strengthening trust in AI systems, especially in the management of digital assets. He also proposes a security model based on human-machine collaboration, where both act as a double-verification system to reduce errors and attacks.
Despite his commitment to local systems, Buterin acknowledges that current AI still depends in many cases on more advanced models hosted in the cloud. To integrate these services without compromising privacy, he proposes the use of technologies such as zero-knowledge proofs, anonymization networks and secure execution environments, which would allow external queries without revealing the user's identity or data.
Another key element of his proposal is reducing dependence on the internet. Self-sovereign AI relies on locally stored knowledge bases, including technical documentation and encyclopedic content, which makes it possible to answer many queries without resorting to external services. This not only improves privacy, but also reduces exposure to attacks and tracking of online activity.
A self-sovereign AI
In short, Buterin's proposal offers a direct alternative to the industry's dominant model, based on centralized platforms and cloud processing. His proposal consists of applying principles such as decentralization, censorship resistance and individual control to artificial intelligence. Although he acknowledges that the technology is not yet fully ready for this shift, he considers that moving toward a self-sovereign AI will be key to preserving privacy in the next decade.