How We Measure

It is unsettling to defend against a capability no one has measured. Each study asks one question in one of two registers: how much AI changes what an attacker can do, and how much it changes what a defender can do.

We run a task to set a baseline, then run it again with AI, holding access and conditions equal. What is left is the difference the AI made, separated from any advantage of access or resources.

Before a study runs, we write down its protocol, endpoints, and analysis plan, and pre-register them. The harness enforces the pre-registration: an altered protocol is refused, a widened study is refused. The question stays fixed before the answer is known.

We run against environments we build and own, so the true state of the target is known by construction. Each study carries a guardrail against its own most likely way of being wrong, and states its limits in the protocol.

For more about how we work, take a look at our Labs.