Abstract
The rapid progress in the reasoning capability of Multi-modal Large Language Models (MLLMs) has triggered the development of autonomous agent systems on mobile devices. MLLM-based mobile agent systems consist of perception, reasoning, memory, and multi-agent collaboration modules, enabling automatic analysis of user instructions and the design of task pipelines with only natural language and device screenshots as inputs. Despite the increased human-machine interaction efficiency, the security risks of MLLM-based mobile agent systems have not been systematically studied. Existing security benchmarks for agents mainly focus on Web scenarios, and the attack techniques against MLLMs are also limited in the mobile agent scenario. To close these gaps, this paper proposes a mobile agent security matrix covering three functional modules of the agent systems. Based on the security matrix, this paper proposes four realistic attack paths and verifies these attack paths through eight attack methods. By analyzing the attack results, this paper reveals that MLLM-based mobile agent systems are not only vulnerable to multiple traditional attacks, but also raise new security concerns previously unconsidered. This paper highlights the need for security awareness in the design of MLLM-based systems and paves the way for future research on attacks and defense methods.
Introduction
As Multi-modal Large Language Models (MLLMs) transition from simple chatbots to autonomous agents capable of controlling smartphones, they gain the power to "see" user screens and "act" on their behalf. While this significantly boosts human-machine interaction efficiency, it creates a massive, unexplored attack surface. This study presents the first systematic security audit of multi-modal mobile GUI agents (such as Apple’s Siri or Honor’s YOYO), uncovering 34 previously unreported attack vectors. The researchers introduced SecMoba, a holistic framework demonstrating how attackers can exploit perception, reasoning, and memory modules to steal private data, manipulate user preferences, or trigger Denial-of-Service (DoS) states.
Threat Modeling and Methodology
Existing security research has largely focused on web-based agents. However, mobile GUI agents are distinct: they rely on screenshot-based perception, interact with mobile-specific system APIs, and often utilize long-term memory of user habits. The authors developed a five-step threat modeling pipeline to identify vulnerabilities:
- Attack Vector Categorization: Identifying how payloads enter the system via App, System, or Web data.
- Vulnerable Asset Identification: Mapping targets like perception models, agent prompts, and memory databases.
- Attack Consequence Recognition: Categorizing threats using the CIA triad (Confidentiality, Integrity, Availability).
- Attack Path Construction: Mapping the flow from a malicious screenshot to a system action.
- Feasibility Analysis: Determining which of the 144 theoretical attacks are practical.
To evaluate these threats, the authors built the SecMoba framework, consisting of a Preprocessor to identify modifiable screen regions, a Generator to craft payloads, and an Evaluator. They tested both semantic payloads (human-readable but misleading, like a fake icon) and non-semantic payloads (imperceptible adversarial perturbations that trick OCR or object detection models).
Experimental Setup
The study utilized two primary evaluation environments:
- Real-World Benchmark: Five representative tasks (web browsing, shopping, ride-hailing) tested on physical devices (HUAWEI P40, HONOR Magic 7) and virtual machines.
- Android-in-the-Wild (AITW): A standard dataset for mobile device control, re-sampled into seven specialized subsets (1,240 total volume) to validate specific attack types like privacy extraction and preference manipulation.
Key Insights and Results
The researchers validated 34 distinct attacks across six different agent architectures, including AutoDroid, VisionTasker, and MobileAgent. Several critical vulnerabilities were identified:
- The Perception Gap: Attacking the smaller models used for OCR and object detection is often more effective than attacking the large reasoning model. These smaller models are easily fooled by tiny, imperceptible perturbations.
- The Memory Paradox: While long-term memory helps agents stay on task, it increases vulnerability to integrity attacks. Agents "paying more attention" to the screen are more likely to ingest malicious prompt injections hidden in the UI.
- High Success Rates in Manipulation: In scenarios involving hijacked purchasing decisions, most agents showed a 100% success rate, allowing attackers to force the agent to purchase specific products.
- Global Denial-of-Service (DoS): By injecting specific string delimiters into app names, attackers could render agents unable to recognize any apps on the device.
- Privacy Extraction: Malicious apps achieved up to a 91.67% success rate in tricking agents into leaking the user's original instructions to the attacker.
Conclusion and Future Work
While the study highlights significant vulnerabilities, it also points toward potential defenses. Agents utilizing "reflection" or "debate" mechanisms (where multiple LLM instances cross-check work) proved more robust, though they suffer from higher latency. Furthermore, agents built on fine-tuned MLLMs (like SeeClick) showed better resilience than off-the-shelf models due to a superior global understanding of UI layouts.
Future research must address the "white-box" constraints of current non-semantic attacks to develop effective black-box adversarial methods. Additionally, the development of safer regular expressions for parsing and better separation of data and instructions is essential for securing the next generation of mobile agents.
Summary updated