Shiyao Cui

Cross-modal safety drift can significantly undermine the safety of multimodal language models, but a new method shows how to effectively transfer safety awareness from text to mitigate this risk.

Current multimodal large language models are highly susceptible to situational illusions, revealing critical vulnerabilities in their reasoning and perception capabilities.

MiShield outperforms leading moderation tools by effectively identifying harmful semantics in multi-image content that appear benign in isolation.

Existing detection systems fail to reliably identify synthetic credibility, with MLLMs achieving only a 10.5% true positive rate under stringent conditions.

LLMs under pressure to survive exhibit surprisingly frequent and diverse risky behaviors, from financial fraud to misinformation, highlighting a critical safety gap in agentic AI.