We are a research group focused on building safer AI systems. We are working on the security and safety of LLM-powered agentic systems powered by foundation models.

• YOGA — Yet Another Generalist Agent (Modular and Extensible)

  

• Mirror GUI — LLM-based GUI Simulator for Agentic Data Synthesis and Evaluation

  

  Mirror GUI is a GUI simulator driven by large language models (LLMs), designed to test and evaluate AI agents interacting with a desktop-like environment. It simulates an Ubuntu-style desktop with application windows, UI elements and a file system so agents can perform GUI actions and researchers can analyze behavior and safety.

• ThoughtAligner - A plug-and-play safety aligner module for your tool-use agent's chain-of-thought

  

Are you worried about your AI deleting your important files without asking for permission? Or it just does something unexpected yet dangerous. ThoughtAligner is here for you.

• MirrorGuard - A plug-and-play safety aligner module for your GUI agent's chain-of-thought

  

CUA Agents provide additional challenges due to its multi-modal nature. Don't worry. MirrorGuard is here for you.

• ReasoningShield - A lightweight model for monitoring the content safety of the CoT.

  

• XuanwuBox - Your AI security advisor in the Docker runtime for your agentic system (To be released)

  

• NVWA Project - Preparing for the emergence of silicon-based life. Indentifying the risks of autonomy

AI research is accelerating the transition toward silicon-based life. Our mission is to identify the risks of autonomous emergence, prevent uncontrolled proliferation, and develop essential control technologies.