Blind Gods and Broken Screens: Architecting a Secure, Intent-Centric Mobile Agent Operating System

Zhenhua Zou, Sheng Guo, Qiuyang Zhan, Lepeng Zhao, Shuo Li, Qi Li, Ke Xu, Mingwei Xu, Zhuotao Liu

TL;DR

Current mobile AI agents are "blind gods" that rely on brittle screen-scraping, making them slow, economically disruptive, and dangerously easy to hack via visual spoofing or prompt injection. This paper proposes Aura, a secure, intent-centric Operating System architecture that replaces unstructured GUI scraping with a structured "Agent Kernel," slashing attack success rates from 40% to 4.4% while delivering a 7x speedup.

What Problem Does This Paper Address?

We are moving from an "App-centric" world to an "Agent-centric" one, where LLM-powered assistants (like Doubao or Honor YOYO) perform tasks across multiple apps. However, most current agents use a "Screen-as-Interface" paradigm. They essentially "look" at the screen, perform OCR, and simulate clicks.

The authors argue this is a fundamental security failure for three reasons:

  1. The Privilege Paradox: To work, these agents need "God Mode" access to the entire screen, exposing sensitive notifications (like 2FA codes) to the LLM.
  2. Structural Vulnerability: Agents are easily fooled by "visual hallucinations" (e.g., a fake "Refresh" button that actually authorizes a payment) and "indirect prompt injections" (malicious instructions hidden in an email or webpage).
  3. Economic Friction: Scraping bypasses the "attention economy" (ads and native UI), leading app developers to block agents entirely.

Methods

The authors propose Aura, a clean-slate architecture built on a Hub-and-Spoke topology. Instead of one monolithic agent with total control, Aura splits responsibilities:

  • System Agent (SA): The "Orchestrator" that parses user intent and creates high-level plans but cannot directly touch app data.
  • App Agents (AA): Sandboxed, domain-specific agents (e.g., a "Booking Agent") that execute tasks via structured APIs rather than pixels.
  • The Agent Kernel: A privileged, TEE-backed (Trusted Execution Environment) security core that mediates all communication.

The Kernel enforces four "Defense Pillars":

  1. Identity Infrastructure: Uses a Global Agent Registry and "Agent Identity Cards" (AIC) to cryptographically verify that an app is who it says it is.
  2. Semantic Firewall: Sanitizes inputs, redacts PII (Personally Identifiable Information), and uses XML-based prompt isolation to prevent malicious data from being interpreted as commands.
  3. Cognitive Integrity: Implements taint-aware memory. It tags data from untrusted sources (like the web) and prevents that data from flowing into "Critical Nodes" (like a bank transfer API) without explicit user consent.
  4. Auditable Execution: Provides a non-deniable, on-device audit trail of every "thought" and action the agent took.

Datasets

The researchers evaluated Aura using MobileSafetyBench, a benchmark designed to test the safety of autonomous mobile agents.

  • They curated a subset of 80 tasks: 35 low-risk functional tasks (e.g., "set an alarm") and 45 high-risk safety tasks (e.g., "forward this private email").
  • Aura was compared against two baselines: Doubao-Standard and Doubao-Pro, which are commercially deployed, system-integrated agents.

Key Insights and Ideas

  • From Pixels to Intents: The "Aha!" moment is that the GUI is the bottleneck. By moving to a structured, agent-native interaction model (APIs), you don't just fix security; you solve the latency problem inherent in processing high-res screenshots.
  • Taint Tracking for LLMs: The authors successfully adapted a classic OS security concept—taint tracking—to the world of LLM context windows. By marking external data as "TAINTED," the system can programmatically intercept an agent before it acts on a malicious "hidden" instruction.
  • The Agent Economy: Aura envisions a shift from the "Attention Economy" (where apps fight for your eyeballs) to an "Agent Economy" (where apps compete to fulfill your intents via standardized APIs).

Key Results

The evaluation showed that Aura significantly outperforms current state-of-the-art GUI-based agents:

  • Utility: Aura achieved a 94.3% Task Success Rate, outperforming Doubao-Pro (77.1%).
  • Security: On high-risk tasks, Aura reduced the Attack Success Rate (ASR) to 4.4%, compared to roughly 40% for the Doubao baselines.
  • Efficiency: Aura is 6.4x to 7.5x faster than GUI-based agents. While Doubao took ~441 seconds on average to complete complex tasks, Aura finished in ~68 seconds by eliminating the "screenshot -> upload -> OCR -> reasoning" loop.

Limitations and Future Work

  • Standardization: Aura requires a massive shift in how apps are built. Developers must expose structured APIs and register with a Global Agent Registry, which requires industry-wide consensus.
  • Hardware Dependencies: The security model relies heavily on TEE and hardware-rooted secure boots, which may not be uniformly implemented across all mobile hardware.
  • Model Hallucinations: While Aura fixes architectural flaws, it cannot entirely prevent the underlying LLM from hallucinating or failing to follow complex logic, though it does provide the guardrails to catch those failures before they cause harm.

Summary updated

Abstract

The evolution of Large Language Models (LLMs) has shifted mobile computing from App-centric interactions to system-level autonomous agents. Current implementations predominantly rely on a "Screen-as-Interface" paradigm, which inherits structural vulnerabilities and conflicts with the mobile ecosystem's economic foundations. In this paper, we conduct a systematic security analysis of state-of-the-art mobile agents using Doubao Mobile Assistant as a representative case. We decompose the threat landscape into four dimensions - Agent Identity, External Interface, Internal Reasoning, and Action Execution - revealing critical flaws such as fake App identity, visual spoofing, indirect prompt injection, and unauthorized privilege escalation stemming from a reliance on unstructured visual data. To address these challenges, we propose Aura, an Agent Universal Runtime Architecture for a clean-slate secure agent OS. Aura replaces brittle GUI scraping with a structured, agent-native interaction model. It adopts a Hub-and-Spoke topology where a privileged System Agent orchestrates intent, sandboxed App Agents execute domain-specific tasks, and the Agent Kernel mediates all communication. The Agent Kernel enforces four defense pillars: (i) cryptographic identity binding via a Global Agent Registry; (ii) semantic input sanitization through a multilayer Semantic Firewall; (iii) cognitive integrity via taint-aware memory and plan-trajectory alignment; and (iv) granular access control with non-deniable auditing. Evaluation on MobileSafetyBench shows that, compared to Doubao, Aura improves low-risk Task Success Rate from roughly 75% to 94.3%, reduces high-risk Attack Success Rate from roughly 40% to 4.4%, and achieves near-order-of-magnitude latency gains. These results demonstrate Aura as a viable, secure alternative to the "Screen-as-Interface" paradigm.
BibTeX
@misc{zou2026blind,
  author = {Zhenhua Zou and Sheng Guo and Qiuyang Zhan and Lepeng Zhao and Shuo Li and Qi Li and Ke Xu and Mingwei Xu and Zhuotao Liu},
  title = {{Blind Gods and Broken Screens: Architecting a Secure, Intent-Centric Mobile Agent Operating System}},
  year = {2026},
  eprint = {2602.10915},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.10915},
}
Figure