ABOUT ANIFOLD

Safer AI.Greater Capabilities.

Anifold builds safer, more capable models and agent systems. We combine AI safety research with training systems engineering to help teams tackle harder problems, operate with greater autonomy, and continuously improve their models while keeping them reliable and under control.

Our perspective

Safety technology.New capability frontiers.

We believe safety is a core capability that can be trained, evaluated, and continuously improved. Learning to recognize risk, handle adversarial conditions, and recover from failure equips models to act reliably in complex environments and take on harder tasks.

Safety research also opens new paths to capability. Through adversarial elicitation and CoT distillation, we uncover capabilities models cannot reach on their own and bring those breakthroughs into post-training. Our goal is for safety and capability to reinforce each other: models become safer as they grow more capable, and more capable as they grow safer.

From research to delivery

Build Safety. Unlock Capability. Evolve Both.

We turn three research and engineering paths into usable environments, data, models, and training services that help customers continuously improve their AI.

Build Safety
Cyber and AI safety Docker environments, an attack knowledge base, and safety data from malicious agent trajectories and harmful content support the training and evaluation of core safety capabilities.
Unlock Capability
Capability unlocking, CoT distillation, and post-training deliver custom models and agents for specific tasks and domains, alongside safety-enhanced open-source models. Services include capability evaluation, red teaming, and safety alignment.
Evolve Both
Our Safe Recursive Intelligence (SRI) platform connects rollouts, evaluation, and training through ready-to-use Docker environments and rollout harnesses, with end-to-end training for customer models to jointly improve safety and capability.

+03 — THE PEOPLE

Built by people who know the frontier

Our work brings together AI safety and alignment, infrastructure, post-training, multi-agent systems, and environments and data.

Xingjun Ma

AI Safety Alignment

Focused on AI safety and alignment, studying how models and agents can remain safe, reliable, and aligned with human intent as their capabilities improve.

Fudan University · Shanghai Innovation Institute (SII)

Mark Jiang

AI infrastructure

Focused on AI infrastructure, building systems that support model training, deployment, and continuous iteration.

University of Melbourne · UCLA · Stanford University

Xiuchuan Ding

Multi-agent systems

Focused on multi-agent systems, studying how agents collaborate, interact, and carry out tasks together.

Fudan University · University of Wisconsin

Jianjian Zhang

Business · GTM

Serial entrepreneur, and co-founder and COO of Dongfangdi. Previously at Baidu, JD.com and Wanda, with nearly 15 years in internet entrepreneurship and product operations, building several businesses from zero to one. Has led teams of more than 100 people, and now leads customer development, delivery and company operations.

Baidu · JD.com · Wanda

Ming Wen

Model post-training

Focused on post-training, using training and feedback to improve model capabilities and behavior.

Fudan University

Yunhao Chen

Environments and data

Focused on environments and data, building interaction environments and data foundations for model training and evaluation.

Fudan University

Start with your task.Train better AI.

Work with us on safety environments and data, custom models and agents for your domain, or infrastructure and training services for continuous improvement. Tell us about your tasks, existing models, and training goals so we can shape the right engagement together.