Alignment failure in deployment
A deployed system is caught pursuing hidden goals, and labs or governments pause frontier development.
Cuts off 6 boxes, 5 of 53 stages · slows 6 · Show on the map →
Hits directly
Signs
Apollo Research finds scheming can be reduced in tests but not eliminated, and models increasingly notice when they are being tested. Forethought recommends commitments to pause if AI research automation passes set thresholds.
What would soften it
Interpretability and monitoring good enough to catch it early and fix it without a pause.