OmniGuard: A unified omni-modal guardrail with deep-thinking reasoning capability

Researchers propose OmniGuard, a family of omni-modal guardrail models with deep-thinking reasoning capability, which can uniformly perform safety protection on text, images, video and audio. Its training dataset contains more than 210K diverse samples, covering unimodal and cross-modal inputs, with each sample annotated with structured safety labels and safety critiques distilled from expert models. Experiments on 15 benchmarks show that OmniGuard has strong effectiveness and generalization capability across a wide range of multimodal safety scenarios.