New benchmark MobileInject evaluates MLLM mobile agent security against multimodal injection attacks

Researchers have introduced MobileInject, a novel benchmark designed to evaluate the security of multimodal large language model (MLLM)-based mobile agents. This framework addresses the emerging threat of adversaries embedding malicious objectives in cross-application mobile content, steering agent behavior through indirect multimodal injection. MobileInject encompasses 11 scenarios, 9 security-risk categories, and 108 executable risk instances, incorporating advanced attack methods like Multi-Role Collaborative Deception and Reasoning-Camouflage Injection. Experiments reveal significant vulnerabilities in mobile agents, often stemming from reasoning hijacking. AI

IMPACT This benchmark could lead to more secure AI agents operating in mobile environments, reducing risks from malicious content injection.

RANK_REASON The item is a research paper detailing a new benchmark for evaluating AI security. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Cross-Modal Consistency Injection
• Hugging Face
• MobileInject
• multimodal large language model
• Multi-Role Collaborative Deception
• Reasoning-Camouflage Injection
• Reasoning Security Diagnosis Framework

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →