We build safer, more capable models and agent systems that tackle harder problems, operate with greater autonomy, and continuously improve while remaining reliable and under control.
We first build core safety capabilities through large-scale training environments. We then use safety technology to uncover capabilities models cannot reach on their own. Finally, we bring safety and capability into a unified RSI framework, redesigning post-training infrastructure and architectures so both continuously evolve and reinforce each other.
Build Safety. Unlock Capability. Evolve Both.
01 · SAFETY ENVIRONMENTS
Build Safety.
Train the safety capability that lets AI go further.
We build large-scale cyber and AI safety environments where agents learn to handle risks, adversarial conditions, and failures. By training safety as a core capability, we create the foundation for agents to take on harder tasks, operate with greater autonomy, and push beyond their current limits.
Safety First. Capability Next.
02 · CAPABILITY UNLOCKING
Unlock Capability.
Use safety technology to uncover capabilities models can't reach on their own.
Through adversarial elicitation, CoT distillation, and boundary jailbreak, we push models beyond what they can achieve on their own—uncovering stronger reasoning and new capabilities, then turning these breakthroughs into training data for even more capable models.
Safety Explores What AI Can Become.
03 · POST-TRAINING INFRA
Evolve Both.
Redesign post-training so safety and capability improve as one.
We build the infrastructure and training architecture for joint recursive self-improvement, where safety and capability continuously learn from and reinforce each other—making models safer as they grow more capable, and more capable as they grow safer.
Safety and Capability. One Evolution.
03 · POST-TRAINING INFRA
Evolve Both.
Redesign post-training so safety and capability improve as one.
We build the infrastructure and training architecture for joint recursive self-improvement, where safety and capability continuously learn from and reinforce each other—making models safer as they grow more capable, and more capable as they grow safer.
Safety and Capability. One Evolution.
ANIFOLD
Build safer AI.Push capability further.
Data × Platform × Safety → Capability
We use safety to expand what AI can do—making agents more capable, autonomous, and adaptive, while keeping them reliable and under control.
Safety → Capability
ANIFOLD · MAKING AI SAFER AND MORE CAPABLE
Build safer AI. Push capability further.
Safety UnlocksCapability
We use safety to expand what AI can do—making agents more capable, autonomous, and adaptive, while keeping them reliable and under control.
Empowering Everyone to Train Safer, More Capable Models.Scroll to view↓
01 · SAFETY ENVIRONMENTS & DATA
Environments and Data
Frontier risk tracking, attack knowledge, and large-scale safety data from malicious agent trajectories and harmful conversations to help teams train and evaluate safety capabilities.
Environments: Containerized tasks and adversarial scenarios for training and evaluation.
Data: Attack knowledge base and large-scale safety data of malicious agent trajectories and harmful content.
Deliverables: Task Environments · Training Datasets · Customization and Evaluation
02 · MODELS & TRAINING SERVICES
Custom Models and Agents
Custom models and agents for specific tasks and domains, with capability unlocking, CoT distillation, and post-training services to make them safer and more capable.
Models and Agents: Custom models and agents for target tasks, and safety-enhanced open-source models.
Training Services: Custom post-training, capability evaluation, red teaming, and safety alignment.
Deliverables: Custom Models · Agent Systems · Evaluation Report
03 · SRI Platform
Safe Recursive Intelligence
An SRI platform connecting environments, rollouts, evaluation, and training to enable AI to become safer as it grows more capable—and more capable as it grows safer.
SRI Infra: Ready-to-use Docker environments and rollout harnesses for training and evaluation.
RSI Training: End-to-end training to jointly improve safety and capability.
