Evolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards

The integration of diverse data modalities into Multi-modal Large Language Models (MLLMs) introduces complex safety challenges beyond those encountered in single-modality systems. This shift necessitates a re-evaluation of existing threat models and safety frameworks, as current approaches are insufficient to address novel risks arising from compromised modality integration, misalignment, and fused safety issues. The evolving landscape requires a systematic understanding of these new threats to develop effective safeguards.

Why it matters

The rapid advancement of Multi-modal Large Language Models (MLLMs) presents both significant opportunities and profound risks. Understanding and mitigating these emerging threats is critical for organizations leveraging or developing MLLMs to ensure system integrity, prevent unintended harmful outcomes, and maintain public trust. Failure to adapt safety protocols to these new complexities could lead to significant operational disruptions, reputational damage, and regulatory scrutiny.

Key insights

• MLLMs integrate heterogeneous modalities, enhancing understanding and reasoning capabilities.
• The architectural shift in MLLMs redefines the machine learning safety landscape.
• Increased model complexity and cross-modal interactions generate novel threats.
• New threats include compromised modality integration, modality misalignment, and fused safety risks.
• Threat modeling for MLLMs must move beyond uni-modal assumptions.
• Existing safety frameworks, rooted in uni-modal learning, are inadequate for MLLMs.
• A systematic analysis of MLLM safety is required due to these evolving challenges.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.07535