Ethereum co-founder Vitalik Buterin shared a new round of updates on privacy-focused AI infrastructure, outlining tools designed to support local AI use, secure messaging, and protected access to Ethereum services.
In an X post on Thursday, Buterin highlighted several open-source projects tied to what he describes as “CROPS AI,” a framework focused on secure, self-sovereign, and privacy-preserving AI systems.
The update builds on a longer essay he released in April about local AI setups and the security risks tied to cloud-based AI agents.
New AI and messaging tools added
Among the projects mentioned was an updated version of Buterin’s messaging-daemon, which now includes alpha Telegram support. The tool is designed to let AI systems interact with messaging apps while restricting high-risk actions through manual confirmation layers.
He also pointed to developments around VoxTerm, a local AI transcription tool that runs without third-party servers, and Lucebox Hub, software aimed at improving the performance of dense AI models such as Qwen 27B.
Buterin additionally referenced a lightweight 2-bit quantized version of DeepSeek v4 capable of running within 90 GB of memory, though he noted performance remains uneven across hardware vendors. According to him, the model currently performs significantly faster on Apple hardware than on AMD systems.
Privacy tools linked to Ethereum access
A major focus of Buterin’s latest comments was the overlap between AI privacy infrastructure and Ethereum access systems. He said zero-knowledge-based API calls designed for private access to remote AI models could also help solve privacy issues around Ethereum RPC reads. The idea would allow users to interact with blockchain infrastructure without exposing identifiable request data to service providers.
Buterin also discussed application-specific AI models tailored for Ethereum development and security research. He cited Leanstral, a model optimized for Lean programming language tasks, as an example of how specialized AI systems can improve software verification and secure coding practices.
“We should have models finetuned for Ethereum-related use cases as well,” he wrote.
April essay focused on local AI security
The latest update follows Buterin’s April essay titled “My self-sovereign / local / private / secure LLM setup.” In that post, he warned against treating current AI systems as secure by default and argued that most mainstream AI ecosystems remain weak on privacy and operational safeguards. He specifically raised concerns about AI agents with unrestricted permissions, malicious prompt injection attacks, data exfiltration risks, and remote cloud models collecting sensitive user information.
To reduce those risks, Buterin described a setup centered on local-first AI models, sandboxed environments, offline knowledge storage, and human approval systems for sensitive actions such as crypto transactions or external messaging. His workflow includes tools like llama-server, bubblewrap sandboxing, and AI agents connected to Ethereum wallets through permission-restricted middleware.
Human confirmation remains central
Buterin argued that AI systems handling financial or messaging functions should not operate autonomously without safeguards. Instead, he proposed combining AI systems with mandatory human confirmation for sensitive tasks. He described the model as a “human + LLM” confirmation framework, where both the user and the AI must approve high-risk actions before execution.
The broader goal, according to Buterin, is to develop AI systems that improve productivity without expanding surveillance risks or exposing users to automated exploits tied to centralized AI infrastructure.
Also Read: Vitalik: Ethereum Has Enough Privacy Narratives as Kohaku SDK Advances
