01
Realistic tasks
Train on security work that requires planning, tool use, and actions across a live environment.
Capability environments / For frontier model companies
License cyber environments for reinforcement learning and model development. Give agents realistic offensive and defensive tasks, repeatable systems to act on, and rewards tied to what they accomplish.
01
Train on security work that requires planning, tool use, and actions across a live environment.
02
Use the resulting system state to determine whether an objective was reached, a threat removed, or a service restored.
03
Start from controlled conditions, vary the scenario, and compare behavior across training runs.
The engagement
The systems, task definitions, starting conditions, and dependencies required for the agreed capability.
Checks that turn verified outcomes into a training signal, with failure cases and limitations documented.
Defined initialization and reset behavior, plus scenario variations suited to your training objective.
An agreed interface to your training stack, documentation, and support scope for adopting the environment.
The training loop
A model acts through the agreed tools.
Actions change the state of a real system.
Independent checks verify the outcome.
The training stack receives the result.
Working with us
Identify the behaviors you want to develop, the agent’s tools, and the systems it needs to interact with.
Agree on task coverage, reward design, deployment requirements, licensing, and integration with your training stack.
Validate the environment and reward behavior, then use training evidence to guide scenario development.
Frontier model companies and teams developing cyber-capable agents. Engagements focus on the model, training stack, and capability objectives you need to support.
Yes. We can scope an environment around an offensive or defensive capability, with explicit tools, starting access, objectives, and reward criteria.
No. Training and evaluation can share infrastructure and grading interfaces, but held-out benchmark tasks, seeds, answers, and evaluation traces stay separate.