Survey Maps Evolving Safety Threats in Multi-modal Large Language Models

A new survey on arXiv (2608.07535) systematically analyzes the evolving safety landscape of multi-modal large language models (MLLMs), which integrate heterogeneous modalities through alignment and fusion. The authors propose a multimodal grounded taxonomy of safety threats, covering adversarial attacks, data poisoning, jailbreaks, and hallucinations. They highlight novel threats arising from increased model complexity and cross-modal interactions, including compromised modality integration, modality misalignment, and fused safety risks. These shifts in threat modeling go beyond uni-modal assumptions, imposing new constraints on safety solutions not captured by existing frameworks. The survey aims to provide a systematic analysis to guide future research and development of safeguards for MLLMs.