Multimodal large language model (MLLM)-based mobile agents are rapidly shifting from conversational assistants to autonomous systems that perceive, reason over, and act on smartphone environments. That growing capability, however, also expands the attack surface, according to a new arXiv paper introducing MobileInject.

The paper proposes MobileInject as an evaluation method for what it calls "reasoning-to-action security." The approach focuses on cross-modal consistency injection, where an attacker introduces conflicting information across different modalities (such as vision and text) to mislead the agent's reasoning and trigger unsafe actions.

Because the abstract is truncated, the full scope of the evaluation results is not available here. Still, the framing makes clear that as mobile agents gain autonomy, ensuring they can resist cross-modal manipulation will be critical to their safe deployment.