Tsofit Zazon

The rapid integration of generative AI technologies, and particularly multimodal models, presents an exciting frontier for tackling complex and lengthy tasks. However, it also opens the door to potential security risks in the cybersecurity landscape. In this talk, I will delve into the security of multimodal large language models (LLMs) by demonstrating various image-based attacks on multimodal systems, illustrating the vulnerabilities these models may possess. Throughout the talk, I will explore the different entry points where an attacker might compromise these models, highlighting the necessity for robust security measures. Then, I will present an approach for safeguarding tasks that are using LLMs by a benchmarking process aimed at determining the robustness of a model for differentiating tasks. This process aids not only in selecting the right model for the right task but is also pivotal in revealing the weak points that could be exploited by adversaries. Moreover, I will address how the benchmark can face future challenges in securing multimodal LLMs for our desired tasks in an ever-changing world.