This paper presents a sandboxed capability model and memory isolation strategy for LLM agents running locally on user devices.
Secure On‑Device Agents for Personal Data
Authors: Your Name
Published: Privacy & AI (PAI) · 2025
Related Papers
Neural Rendering Techniques for Real‑Time UX
We explore efficient neural rendering pipelines enabling interactive user experiences on the web with constrained compute and latency.
Analysis of Undefined Behavior in C: The Pedagogical Issues of Using Multiple Pre- and Post-Increment Operators Across Compilers
This paper provides an in-depth analysis of why expressions in C that combine multiple pre- and post-increment operators on a single variable exhibit unpredictable and compiler-dependent behavior. The analysis shows that this phenomenon is a direct consequence of unsequenced modifications, a specific form of undefined behavior sanctioned by the C language standard. The historical rationale behind this design choice, intended to enable aggressive compiler optimizations, is explored. Through case studies of divergent compiler outcomes, the practical risks of relying on such expressions are illustrated. The paper concludes that these expressions are pedagogically unsound and should be discouraged in foundational C curricula in favor of clarity, defensive coding, and an understanding of the language's formal rules.