Reproduce “Blind Gods and Broken Screens: Architecting a Secure, Intent-Centric Mobile Agent Operating System” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhenhua Zou, Sheng Guo, Qiuyang Zhan, Lepeng Zhao, Shuo Li, Qi Li, Ke Xu, Mingwei Xu and 1 more
Aura, a secure mobile agent operating system architecture from Tsinghua University, re-architects AI agents from interacting via visual interfaces to a structured, intent-centric model, addressing inherent security flaws and efficiency bottlenecks. It achieved a 94.3% task success rate and reduced the attack success rate to 4.4%, while accelerating task completion by 6.4x compared to existing GUI-based agents.
Key insights
- A systematic security analysis of production mobile agents like Doubao Mobile Assistant confirms that the "Screen-as-Interface" model introduces fundamental vulnerabilities across agent identity, perception, cognition, and execution.
- A "clean-slate" operating system architecture with kernel-level security enforcement is essential to address the root causes of these vulnerabilities, rather than relying on superficial mitigations.
- Transitioning from unstructured visual data to structured, intent-centric protocols and APIs provides substantial gains in both security by enabling deterministic policy enforcement and efficiency by eliminating visual processing overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.