Goodfire

Reading a model's internals instead of its output cut the cost of watching for cyber misuse by about thirty times, Goodfire says

A lightweight probe reads activations at every turn and escalates only what looks wrong, which drops judge cost from $3,166 to $65 per million turns.

Three companies that host and serve open models will try to write a safety standard for them

The target is abliteration, the practice of surgically removing a model's refusals. The three say safety has to be built into how open models are trained and served, not bolted on.