Xuhong Wang

This page uses posts to introduce several representative works from our team, including papers, technical reports, system releases, and research summaries.

Security as a Service: Releasing Mokai, a 150+ Tool Full-Stack Security Toolbox for Agents

Security as a Service: Releasing Mokai, a 150+ Tool Full-Stack Security Toolbox for Agents

Releasing Mokai, a full-stack agent security toolbox that bundles 150+ tools across risk simulation, trusted data, and evolutionary defense for flexible deployment and real-world adoption.

Apr 27, 2026

Navimaster:The First Unified Navigation Model Across Digital and Physical Worlds, with Minecraft Support

Navigation Agent

Navimaster:The First Unified Navigation Model Across Digital and Physical Worlds, with Minecraft Support

NaviMaster unifies GUI navigation and embodied navigation within a single training framework. Through a shared trajectory formulation, unified reinforcement learning, and distance-aware dense rewards, it improves cross-task generalization, training efficiency, and spatial precision at the same time.

Xuhong Wang

• Mar 18, 2026

Small Models Read, Large Models Reason: Releasing DRIFT for Efficient and Jailbreak-Resistant Reasoning

Small Models Read, Large Models Reason: Releasing DRIFT for Efficient and Jailbreak-Resistant Reasoning

DRIFT decouples long-context reading from reasoning, using a small model to extract dense task-relevant knowledge and a large model to reason more efficiently and robustly.

Mar 14, 2026

SafeVerse: Building a Safe and Trustworthy Digital Twin Arena for Embodied AI

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.

Xuhong Wang

• Feb 10, 2026

BioBridge: Letting LLMs Truly Understand Proteins Without Sacrificing General Ability

BioBridge: Letting LLMs Truly Understand Proteins Without Sacrificing General Ability

BioBridge combines domain-incremental continual pretraining, protein-language semantic alignment, and end-to-end multitask modeling to connect the reasoning strength of general LLMs with the domain expertise of protein models, reaching near-specialist performance on real biological tasks while preserving general capability.

Xuhong Wang

• Dec 15, 2025

Embodied AI Is Racing Ahead, but Safety Lags Behind: Releasing a First Roadmap for Safe and Trustworthy EAI

Embodied AI Is Racing Ahead, but Safety Lags Behind: Releasing a First Roadmap for Safe and Trustworthy EAI

Releasing a roadmap for Safe and Trustworthy Embodied AI, defining the field with ten core principles and an L1-L5 maturity model for embodied AI safety.

Sep 16, 2025

SafeWork-R1: Co-Evolving Intelligence and Safety under the AI-45 Degree Law

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.

Xuhong Wang

• Jul 12, 2025