Running large language models locally is often seen as a privacy safeguard: prompts never leave the device, unlike cloud-hosted inference. A new paper on arXiv challenges this assumption by asking whether local inference alone is enough to keep user data confidential.

The study focuses on "confidentiality boundary failures" in consumer LLM serving systems. The authors appear to probe where the local processing boundary can be crossed, potentially exposing prompts or outputs despite the model running on-device.

The abstract is truncated, so the specific failure modes are not detailed in the source. However, the paper's title—"The Illusion of Local Privacy"—suggests the answer may be that local serving is not automatically private, and that the boundaries around local inference deserve closer scrutiny.