Definition
A test designed to measure whether an AI model can actually carry out a harmful task on its own, such as finding and exploiting a software vulnerability. A model scoring near the ceiling on such a benchmark signals the capability threshold for autonomous cyber offense has been crossed.
Why it matters
These capability readings drive risk and liability decisions — what to allow, what to license, and who is responsible if an agent misuses the skill — so boards need to read them, not just headlines.