Running large language models on edge devices is often presented as a way to keep sensitive data off the cloud. A new preprint on arXiv challenges that assumption by showing that local inference can still leak information through physical side channels.
The researchers demonstrate attacks that recover both output tokens and user prompts while the model is running locally. This suggests that an attacker with physical access to a device could potentially bypass the privacy benefits of on-device processing.
The finding does not mean edge LLMs are useless, but it does indicate that privacy claims based solely on local execution are incomplete. Physical security and side-channel resistance need to be part of the design conversation.