Distributed Implicit Harm: A Compositional Safety Blind Spot in MLLM-Based Video Moderation

A new arXiv study (2609.00206v1) reveals that multimodal large language models (MLLMs) used for video moderation fail to detect 'Distributed Implicit Harm' (DIH), where individually benign video components combine to convey harmful meaning. The researchers developed a multi-agent synthesis framework to create a dataset of over 9,000 DIH videos and benchmarked over 30 MLLMs, including frontier proprietary and open-source models, finding consistent detection deficits across both temporal and cross-modal harm, with failures persisting even in the strongest models and on real-world social media videos.

arXiv:2609.00206v1 Announce Type: new Abstract: Despite their growing use in video moderation, multimodal large language models MLLMs exhibit a compositional safety blind spot: videos composed of seemingly benign components can convey harmful meaning when interpreted as a whole. We refer to this phenomenon as Distributed Implicit Harm DIH , where harm arises from relations among components distributed along a decomposition axis of the video, rather than from any single explicit cue. Among many possible axes, we study two representative cases: temporally distributed harm across visual segments DIH-T and cross-modal harm between audio and visual streams DIH-M . Studying and mitigating DIH at scale requires data that is difficult to collect: such videos lack compositional harm annotations, evade retrieval based on local visual cues, keywords, or single-modality signals, and are consequently absent from existing safety datasets. To bridge this gap, we develop a multi-agent synthesis framework that composes individually benign components into harmful scenarios and generates diverse DIH videos with explicit reasoning annotations, yielding a dataset of over 9,000 videos spanning visual-only and audio-visual settings. Benchmarking over 30 MLLMs spanning frontier proprietary models and leading open-source systems reveals substantial and consistent deficits in detecting both DIH-T and DIH-M. Notably, this failure persists even among the strongest frontier models: they often correctly assess individual components in isolation but fail to recognize the harmful meaning that emerges from their composition. We further evaluate these models on a manually collected set of real-world DIH videos from social media and observe the same failure mode, highlighting DIH as a practical and underexplored challenge for video moderation.