Multimodal Safety and Robustness

As generative image models become increasingly capable, securing them against misuse and unexpected behaviors is vital. Our research addresses the core security challenges of visual AI, ranging from adversarial prompt manipulation and cross-modal jailbreaks to synthetic misinterpretation and algorithmic bias. By developing resilient safety guardrails and robust alignment frameworks, we aim to ensure image generation technologies remain reliable, ethical, and secure against real-world vulnerabilities.