Number 3 of 12 · November 14, 2026

MobileProbe: Adaptive Security Evaluation of Mobile Agents via Trajectory-Aware Injection

MobileProbe tests phone-operating SI agents by adapting attacks to the route each agent takes, and the authors report it beats fixed, hand-written tests at finding weaknesses.

What they found

The authors study mobile GUI agents, which read a phone screen and use apps for a user. An attacker can hide malicious instructions in ordinary screen content to redirect the agent. This is called an environmental injection attack. Current benchmarks pin hand-written attacks to one step of one task. The authors argue this misses agents that take a different route and never see the attack, so they are wrongly marked safe. MobileProbe finds spots where an attacker can control content, finds weak decision points along the agent's route, and refines attack text through an evolutionary loop filtered by a surrogate model. The authors built a benchmark of 83 long-horizon attack instances across 12 mobile apps, each tagged with a Confidentiality, Integrity, or Availability goal. On three agent architectures, MobileProbe got higher attack success rates than human-written, zero-shot LLM, and prior automated red-teaming baselines. The abstract gives no specific success numbers.

Why it matters

SI agents that act on phones can send messages and buy things, so a hidden instruction on screen can cause real harm. If fixed benchmarks undercount the risk, builders may think their agents are safer than they are. The authors present MobileProbe as a pre-deployment audit tool to find weaknesses and guide hardening.

How it changes the race

It raises the bar for security testing of mobile SI agents, since a fixed benchmark pass may no longer be enough to show an agent is safe. It is an evaluation tool, not a new capability, and it may add testing cost for teams shipping these agents.

Safety and control

The paper bears directly on security. It shows an automated way to craft attacks that redirect mobile agents, which helps defenders audit systems but could also be used by attackers. The abstract states the goal as auditing and hardening.

The abstract, by the authors

Mobile GUI agents read the screen of a phone and operate apps on a user's behalf to complete multi-step tasks such as online shopping or messaging. Because the agent acts on whatever the screen shows, an attacker can embed malicious instructions into ordinary UI content and redirect its actions. We refer to this threat as environmental injection attack (EIA). Existing benchmarks measure this risk by pinning hand-crafted payloads to a fixed step of a fixed task. This approach faces two limitations. First, each payload is bound to one task in one app and does not transfer without manual redesign. Second, agents given the same task may follow different routes, so an agent that never reaches the injected page is recorded as safe without being tested. This insufficient test coverage can cause fixed-point benchmarks to underestimate the security risk of deployed mobile agents. In this work, we propose MobileProbe, a trajectory-aware red-teaming framework that evaluates mobile agents adaptively. MobileProbe discovers attacker-controllable content surfaces through automated pentesting, localizes vulnerable decision points along the agent's trajectory, and iteratively refines context-aware payloads through a surrogate-filtered evolutionary loop. We construct a benchmark of 83 long-horizon attack instances spanning 12 mobile applications. Each is annotated with a Confidentiality, Integrity, or Availability (CIA) attack goal to cover the principal categories of security threats facing mobile agents. On three representative agent architectures, MobileProbe attains higher attack success rates than human-written, zero-shot LLM, and prior automated red-teaming baselines adapted to the mobile domain. MobileProbe can serve as a pre-deployment auditing tool to systematically uncover vulnerabilities and inform security hardening of mobile agents.

Quoted from arXiv:2611.16693. The research and its words belong to its authors.

The authors

Yingbin Jin, Ji'an Wang, HaoYang Chai, Hanjun Luo, Li Hu, Xinfeng Li, Huadi Zheng, PeiZhao Hu, Qingqing Ye, Haibo Hu

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.