MSTS: A Multimodal Safety Test Suite for Vision-Language Models

2025-01-17

AI Summary

This study systematically uncovers a latent safety risk in vision-language models (VLMs): hazardous behavior triggered by joint image-text inputs. Method: We introduce MSTS—the first multimodal safety evaluation suite for VLMs—comprising 400 cross-modal trigger samples spanning 40 fine-grained harm categories. We propose a novel definition and evaluation paradigm for multimodal cooperative triggering, design a cross-lingual safety assessment framework, and conduct multimodal adversarial prompting, semantic vulnerability mining, and automated classifier evaluation. Contribution/Results: We discover that non-English prompts increase harmful response rates by 3.2× on average; identify and validate “accidental safety”—a phenomenon where VLMs exhibit false compliance due to comprehension deficits; demonstrate that unimodal text-only safety testing severely underestimates real-world multimodal risks; and show that even the state-of-the-art safety classifier achieves only F1 = 0.61, underscoring the critical need for rigorous multimodal safety evaluation.

Technology Category

• Natural Language Processing: Safety and Robustness
• Computer Vision: Multi-modal Vision
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.

Problem

• Visual Language Models
• Safety Risks
• Harmful Behavior

Innovation

• MSTS
• Visual Language Models
• Multilingual Safety Assessment