Vitalik: AI should not have free access to personal data, local priority+sandbox+human-machine dual confirmation is the correct solution Ethereum co-founder Vitalik Buterin shares his exploration of security and privacy for personal use of AI. Vitalik pointed out that current AI is generally lacking in privacy and security, including locally open-source AI. Typical issues include: agents can modify critical settings without human confirmation, malicious external inputs can easily take over user instances, and some skills contain malicious instructions. What is his plan? -Local priority: All LLM inference and files run locally, without relying on the cloud -Sandbox isolation: Strictly restrict LLM's system access permissions through bubblewrap -Human machine dual confirmation: LLM can only read messages and send messages to itself, and sending messages to others must be confirmed manually In terms of hardware, he tested the NVIDIA 5090 laptop and AMD Ryzen AI Max Pro, running the Qwen3.5:35B model using the NixOS system and pi proxy framework. Vitalik believes that the failure modes of humans and LLM are different, and the dual confirmation mechanism combining the two is more reliable than relying solely on either party. He also called for the construction of a multi-layered defense system, including zero knowledge API calls, hybrid networks, TEE inference, and input cleansing, and suggested transforming every paid API into ZK-API. If developed properly, AI can create a stronger future of privacy and security. "