arXiv 2026 · Preprint

AV-SafetyBench

A Safety Benchmark for Text-to-Audio-Video Generation

Suah Choi Tae-Young Lee Gyeong-Moon Park†

VGI LAB, Korea University

†Corresponding author

Grid of 13 risk categories under four color-coded axes, each with one representative generated sample and the harm-relevant utterance.
Overview AV-SafetyBench overview. 13 risk categories under four color-coded axes (legend), one representative generated sample per category with the harm-relevant utterance beneath. Rose overlays flag the harmful cue and green overlays the context cues that support it. Graphic content is moderately masked.

Abstract

The first safety benchmark built for text-to-audio-video generation

Recent text-to-audio-video (T2AV) models jointly generate video, speech, sound effects, and ambience from a single text prompt. This capability poses new challenges for safety evaluation, as unsafe content may be conveyed through the audio track or arise only when the visual and audio tracks are interpreted jointly. Existing safety benchmarks largely focus on either generated video or generated audio in isolation and are therefore not designed to capture these risks. To close this gap, we introduce AV-SafetyBench, the first safety benchmark developed specifically for T2AV generation. AV-SafetyBench comprises a four-axis, 13-category taxonomy and 5,200 manually reviewed prompts that specify visual scenes, speech, and non-speech audio. Our evaluation protocol assesses each output under three views: Full-AV, Video-Only, and Audio-Only. It then uses the Video-Only and Audio-Only judgments to assign Full-AV unsafe outputs to one of four risk sources: Video-Only, Audio-Only, AV-Both, or AV-Joint. We evaluate five open-source T2AV models and validate the automated Full-AV judgments against human annotations. Across the five models, Full-AV Unsafe Rates range from 25.1% to 49.4%. Beyond these aggregate rates, risk-source analysis reveals that, for four of the five models, Audio-Only and AV-Joint cases—unsafe outputs missed by video-only evaluation—account for 41.6–48.3% of Full-AV unsafe outputs for which a risk source could be assigned. In the Cross-Modal Harm Emergence category, AV-Joint accounts for 87.5% of unsafe outputs with an assigned risk source. Together, these findings demonstrate the value of AV-SafetyBench for evaluating T2AV safety across the visual and audio modalities and their interaction.

Motivation

Video-only evaluation is insufficient for T2AV generation

Unsafe meaning may arise from video, audio, or their interaction.

Four Full-AV unsafe outputs, one per view-based risk source class, with per-view unsafe and safe judgments.
(a) Unsafe Generations by Risk Source
Full-AV Unsafe Rate of five T2AV models on the transferred T2VSafetyBench tinyset and the AV-SafetyBench tinyset.
(b) Full-AV Unsafe Rate by Model
Figure 1 Safety risks in T2AV outputs are not confined to the visual track. (a) Four Full-AV unsafe outputs, one per view-based source class, with red and green badges giving the per-view unsafe and safe judgments; graphic content is moderately masked. (b) Full-AV Unsafe Rate across five T2AV models under the same judge, on the transferred T2VSafetyBench (T2VSB) and AV-SafetyBench (Ours) tinysets.

AV-SafetyBench

Data construction and three-view evaluation

Three-part overview: taxonomy grounding, prompt construction pipeline, and three-view judging.
Figure 3 Overview of AV-SafetyBench data construction and evaluation. (a) 13 risk categories grounded in provider usage policies and prior work. (b) Seed prompts are validated, LLM-expanded, and manually reviewed. (c) Each generated clip is judged under three views in a single pass.

Results

Full-AV Unsafe Rate and where the risk comes from

Examples

Example generations by risk category

These samples contain unsafe audiovisual content. Each clip and its prompt stay hidden until you choose to view them; videos in categories 1–4 are blurred.

News

  • Project page is live. Code and dataset release coming soon.
  • AV-SafetyBench paper released on arXiv.

Citation

BibTeX

@article{choi2026avsafetybench,
  title   = {AV-SafetyBench: A Safety Benchmark for Text-to-Audio-Video Generation},
  author  = {Choi, Suah and Lee, Tae-Young and Park, Gyeong-Moon},
  journal = {arXiv preprint arXiv:2609.06991},
  year    = {2026}
}