MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection
MobileInject benchmarks MLLM mobile agents against multimodal injection across 108 risk instances, finding broad vulnerability.
MobileInject is a benchmark for multimodal injection against MLLM-based mobile agents that perceive, reason, and act on smartphones. It covers 11 scenarios, nine security-risk categories, and 108 executable risk instances, and introduces Cross-Modal Consistency Injection, Multi-Role Collaborative Deception, and Reasoning-Camouflage Injection. Experiments find broad vulnerability, with many failures caused by reasoning hijacking at early stages. A Reasoning Security Diagnosis Framework is included for fine-grained failure analysis.
- Benchmark covers 11 scenarios and 108 executable risk instances.
- Cross-modal consistency injection exploits agreement across modalities.
- Added attacks target trust manipulation and reasoning hijacking.
- Many failures begin with early-stage reasoning hijacking.
From the original article
1 min read at arxiv.orgMultimodal large language model (MLLM)-based mobile agents are rapidly evolving from conversational assistants into autonomous systems that perceive, reason over, and act on smartphone environments. This capability also exposes a new security risk: adversaries can embed malicious objectives in cross-application mobile content and steer agent behavior through indirect multimodal injection.
By Songping Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.