JPRO: An Automated Multimodal Jailbreak Method Based on a Multi-Agent Collaboration Framework

The researchers propose a multi-agent collaboration framework, JPRO, which can automate jailbreaking of vision-language models (VLMs), achieving an attack success rate exceeding 60% on multiple advanced VLMs including GPT-4o, significantly outperforming existing methods. The framework consists of four specialized agents working collaboratively, together with two core modules, Tactic-Driven Seed Generation and an Adaptive Optimization Loop, to generate effective and diverse attack samples. The paper states that existing jailbreak methods either require white-box access, limiting practicality, or rely on manually designed patterns, resulting in poor sample diversity and scalability; JPRO, as a black-box attack method, overcomes these shortcomings. The authors state that JPRO both reveals critical security vulnerabilities in multimodal models and provides a reference for evaluating and enhancing VLM robustness.