Multimodal large language model (MLLM) based mobile agents are rapidly gaining autonomy in perception, planning, and closed-loop execution within dynamic mobile environments. This autonomy, however, also broadens the attack surface and amplifies the risks of malicious exploitation, making security evaluation a pressing concern.

To address this, the paper introduces MobileJailBench, a benchmark specifically designed for jailbreak testing of mobile agents. It provides a structured way to assess how vulnerable these agents are to adversarial prompts that could lead to harmful or unintended actions.

The significance lies in its systematic approach: as mobile agents become more independent, understanding their failure modes under attack is critical. MobileJailBench offers a foundation for comparing defenses and guiding safer deployment of autonomous agents in real-world settings.