Multimodal Large Language Models (MLLMs) operating within multi-agent environments are susceptible to a sophisticated attack known as "infectious jailbreak." This vulnerability is triggered by the introduction of a single adversarial image into the memory of one agent, which can rapidly propagate harmful behaviors across nearly all agents in the system. This propagation occurs exponentially through agent-to-agent interactions, facilitated by pairwise communication, without requiring further intervention from the attacker. The adversarial image effectively functions as a digital "virus," compromising the integrity of the entire network. Mitigation steps: **For AI Developers:** * Restrict the degree and nature of communication between agents to minimize image or information exchange opportunities. * Design and implement active recovery mechanisms to restore infected agents to a safe state efficiently. **For Model Trainers/Fine-tuners:** * Implement periodic sanitization and review processes for agents' memory banks to remove adversarial content. * Develop robust detection methods to identify and flag adversarial images or text within the agent's memory, distinguishing between benign and malicious content. * Investigate and enhance MLLM model architectures to reduce susceptibility to infectious jailbreak attacks, focusing on addressing fundamental vulnerabilities.