Intelligence is becoming agentic.
AI systems are moving beyond generating answers. They are beginning to plan, decide, and act.
Building the Safety Infrastructure for the Agentic Era
Our work spans safety evaluation, safety data generation, foundation model enhancement, runtime defense, and agent governance.
Why We Exist
Agents will operate in the real world.
They will use tools, access systems, collaborate with other agents, and interact with physical environments.
Safety must evolve with intelligence.
As AI becomes more capable and autonomous, safety must become a foundational capability — not an afterthought.
The Challenge
Safety challenges evolve as intelligence becomes more capable.
Our Approach
We address AI safety through three core capabilities that work together to build a complete safety infrastructure for intelligent systems.
Learn MoreRuntime Defense
Protect agents during real-world execution with real-time detection and intervention.
Safety Data Generation
Generate diverse, high-quality safety data at scale to improve coverage and robustness.
Foundation Model Enhancement
Strengthen base models with safety at the core through alignment and capability improvement.
Our Team
Researchers, engineers, and entrepreneurs from leading universities and technology companies around the world.
Industry Experience
OpenAI
Google
Microsoft
ByteDance
Alibaba
...
Academic Background
Fudan
Tsinghua
Stanford
UCLA
USC
...
Our Long-Term Vision
AI safety is a multi-decade challenge. Our roadmap is organized around three stages.
Agent Safety
Building the core safety infrastructure required to evaluate, secure, and govern AI agents.
Safe Agent-Powered Devices
Extending safety capabilities to next-generation intelligent devices and physical systems.
Embodied AI Safety
Developing safety architectures for robots and embodied intelligent systems operating in the physical world.