Definition
A specific danger level — such as the ability to meaningfully help someone create a weapon — that AI labs test their models against before release, and try to suppress with safety training ('mitigation'). A frontier lab recently acknowledged, for the first time publicly, that it could not fully rule out one of its shipped models crossing this threshold before mitigations were applied.
Why it matters
This is a watershed moment in AI safety transparency: it shows that even leading labs cannot always prove their safety measures work before a model reaches the public, which is central to debates over mandatory pre-release evaluation.