Definition
This is the point at which a frontier AI model's ability to find, exploit, or automate cyberattacks becomes so advanced that developers and governments treat it as a distinct, higher-risk tier requiring special safeguards before release. It is one of several 'capability thresholds' AI labs use in their safety frameworks to decide when extra caution or delayed release is warranted.
Why it matters
Crossing this threshold means a model could meaningfully help even unsophisticated attackers conduct serious cyberattacks — boards overseeing AI vendors or deployments should ask whether models they use have been evaluated against thresholds like this and what mitigations were applied.