Safety updates rarely dominate mainstream attention, but they often reveal where labs think the capability frontier is heading. Google DeepMind’s latest Frontier Safety Framework revision adds a formal harmful-manipulation category and new tracked capability levels intended to catch risks earlier.
That matters because the frontier conversation is shifting from static policy language to operational thresholds tied to deployment decisions. The update also broadens how Google thinks about internal deployment risk, not just public launches.
For readers tracking trends, this is a useful reminder that governance is becoming part of product design. The labs are not only shipping stronger models; they are building internal machinery to decide when those models are safe enough to ship.