FlowGuard: catching multimodal AI attacks by checking whether text and image reasoning agree

FlowGuard defends multimodal AI models by checking whether their text-only and vision-only reasoning stay consistent when fused, catching attacks hidden across modalities.