SafeVerse: Building a Safe and Trustworthy Digital Twin Arena for Embodied AI
SafeVerse turns ordinary videos into interactive, physics-aware 3D twin environments within minutes, then extends them with attack-driven scene editing and online agent evolution to form a full reconstruction-defense-evolution loop for embodied safety.
Feb 10, 2026
SafeWork-R1: Co-Evolving Intelligence and Safety under the AI-45 Degree Law
SafeWork-R1 is a multimodal reasoning model built to improve safety and general capability together. Powered by the SafeLadder framework, it substantially outperforms its base model on safety benchmarks while remaining competitive on general reasoning and multimodal tasks.
Jul 12, 2025