Adversarial Attacks on Multi-Modal LLMs: How Image, Audio, and Video Inputs Get Weaponized and How to Defend Against It

NeuralStack | MS — AI Security For most of the past three years, the AI security community's attention has been fixed on text: prompt injection, jailbreak templates, token-level adversarial suffixes.