When Helpers Become Hazards: A Benchmark for Analyzing Multimodal LLM-Powered Safety in Daily Life

TL;DR

This work introduces SaLAD, a 2,013-sample multimodal safety benchmark designed to probe how MLLMs recognize and respond to everyday hazard risks embedded in joint image-text contexts. Using a safety-warning-based evaluation, the study reveals that even top models achieve only 57.2% accuracy on unsafe cases and average around 30.65% overall, with safety alignment methods offering limited improvements. The benchmark emphasizes authentic visuals, no VSIL leakage, and realistic risk scenarios across 10 categories, demanding strong cross-modal reasoning beyond text-only cues. Findings indicate a substantial gap between possessing safety knowledge and applying it multimodally, underscoring the need for advanced cross-modal safety alignment and more robust evaluation frameworks for real-world AI assistants.

Abstract

As Multimodal Large Language Models (MLLMs) become an indispensable assistant in human life, the unsafe content generated by MLLMs poses a danger to human behavior, perpetually overhanging human society like a sword of Damocles. To investigate and evaluate the safety impact of MLLMs responses on human behavior in daily life, we introduce SaLAD, a multimodal safety benchmark which contains 2,013 real-world image-text samples across 10 common categories, with a balanced design covering both unsafe scenarios and cases of oversensitivity. It emphasizes realistic risk exposure, authentic visual inputs, and fine-grained cross-modal reasoning, ensuring that safety risks cannot be inferred from text alone. We further propose a safety-warning-based evaluation framework that encourages models to provide clear and informative safety warnings, rather than generic refusals. Results on 18 MLLMs demonstrate that the top-performing models achieve a safe response rate of only 57.2% on unsafe queries. Moreover, even popular safety alignment methods limit effectiveness of the models in our scenario, revealing the vulnerabilities of current MLLMs in identifying dangerous behaviors in daily life. Our dataset is available at https://github.com/xinyuelou/SaLAD.

Figures (20)

• Figure 1: Examples of existing multimodal safety benchmarks and SaLAD.
• Figure 2: The performance gap of safety defense methods between SIUO and SaLAD.
• Figure 3: Safety taxonomy of SaLAD.
• Figure 4: Unsafe examples of SaLAD, the safe subset is provided in Appendix \ref{['app:example']}. "Q" represents the input query, and "W" denotes the corresponding safety warning.
• Figure 5: Overview of the three-step construction pipeline. The green block represents the benchmark we construct.