Privacy Boundaries in Local AI Execution: Technical Report
This report analyzes data flows in Interaction Task Mode from a system-architecture perspective: which data stays local, which must traverse the network, and the privacy risk profile of each category.
Note: content sent in the LLM context window is processed by third-party model providers. Avoid placing highly sensitive credentials directly in task context.
Key findings: user file contents, intermediate tool execution state, and local codebase contents are architecturally isolated within the local VM and never transmitted through Lumirain servers.
Residual risk: file content fragments mixed into the model context window are transmitted to third-party LLM providers via API calls. Mitigations: explicit context-control UI; minimum-necessary injection principle.