Intelligence Under Authority

The central conviction behind this lab is that cybersecurity is entering a new phase. Models will soon be able not merely to describe vulnerabilities, but to investigate systems, operate tools, and pursue technical objectives over long periods. That capability could make defenders dramatically more effective. It could also magnify mistakes, weak authorization, and misuse.

We think both truths have to be held at once. It is neither responsible to ignore the risks nor useful to treat them as inevitable. Our work is to build the technical path between those extremes: open cyber models trained on grounded evidence, tested in verifiable environments, and governed by controls outside the model itself.

The goal is intelligence that expands the defender's reach without escaping the defender's authority.

Built carefully, this technology can make secure systems easier to create, investigate, and maintain—and powerful cyber capability more accountable.