We’re a cybersecurity AI research company developing large language models to find vulnerabilities in software.
Cybersecurity, chess, and Go share something important: strategies can be tested and successful outcomes can be verified. Inspired by AlphaGo, we’re developing models that learn from those outcomes.
Pretraining has given us capable general-purpose models. We believe the next breakthrough will come from reinforcement learning: training models to earn rewards for solving real problems, beyond predicting the next token.
The model is the moat. An agent built entirely on another company’s model leaves its core capability in someone else’s hands. We’re developing our own models and training systems.
We’re finding vulnerabilities in software with our autonomous pentesting agent. Our models will train on a broad range of cybersecurity data and tasks, including verified findings and the attempts behind them.