MobileInject
The first evaluation benchmark targeting multimodal injection attacks and reasoning-to-action security risks for mobile agents. It provides a high-fidelity interactive mobile environment containing multiple scenario categories, security risk categories, and executable risk instances.
Timeline (last 90 days)
MobileInject is the first evaluation benchmark targeting multimodal injection attacks and 'reasoning-to-action' security risks in mobile agents.
MobileInject provides a high-fidelity interactive mobile evaluation environment containing 11 scenario categories, 9 security risk categories and 108 executable risk instances.
MobileInject proposes Cross-Modal Consistency Injection, whose core idea is to have malicious instructions corroborate each other across multiple modalities to lower the agent's vigilance.
MobileInject's accompanying Reasoning Security Diagnosis Framework is used to perform fine-grained attribution of failure cases, determining whether the problem lies in perception, reasoning or action.
Systematic experiments show that mobile agents are generally vulnerable to multimodal injection attacks, and that a large number of failures stem from hijacking early in reasoning rather than the final click action.
Information on the release of MobileInject code is given in the open science section of the paper's appendix.
All Facts (6)
MobileInject is the first evaluation benchmark targeting multimodal injection attacks and 'reasoning-to-action' security risks in mobile agents.
50%UnverifiedMobileInject provides a high-fidelity interactive mobile evaluation environment containing 11 scenario categories, 9 security risk categories and 108 executable risk instances.
50%UnverifiedMobileInject proposes Cross-Modal Consistency Injection, whose core idea is to have malicious instructions corroborate each other across multiple modalities to lower the agent's vigilance.
50%UnverifiedMobileInject's accompanying Reasoning Security Diagnosis Framework is used to perform fine-grained attribution of failure cases, determining whether the problem lies in perception, reasoning or action.
50%UnverifiedSystematic experiments show that mobile agents are generally vulnerable to multimodal injection attacks, and that a large number of failures stem from hijacking early in reasoning rather than the final click action.
50%UnverifiedInformation on the release of MobileInject code is given in the open science section of the paper's appendix.
50%