New Benchmark Addresses Multimodal AI Audio-Video Safety Risks

Audio-video generation is rapidly evolving beyond simple text prompts to incorporate multimodal conditioning, where text, images, audio, and video inputs jointly influence the output. This shift fundamentally alters safety evaluation, as harmful intent might not be present in any single input but could emerge from the interaction of otherwise benign or weakly harmful conditions across modalities and time. Existing safety benchmarks are largely prompt-centric or tied to fixed interfaces, making it difficult to systematically study these complex compositional risks. To fill this gap, Multi2AV-Safety has been introduced as the first safety benchmark specifically designed for multimodal-to-audio-video generation. It covers all 11 non-singleton conditioning configurations (Text, Image, Audio, Video combinations) and includes 11,024 attack instances. Evaluation using Multi2AV-Safety reveals systematic vulnerabilities in representative multimodal safety guards. These guards exhibit two main failure modes: harmful semantics can arise from the combination of individually benign inputs, and explicit harmful cues become harder to detect when mixed with benign multimodal context. These findings highlight "compositional risk perception" as a critical capability gap, indicating that current safety guards struggle to reliably integrate safety evidence across different modalities and over time, even when all inputs are observable. The dataset will be publicly released in October 2026.