■Controls on dangerous artificial intelligence (AI) capabilities can weaken on their own, even when policy stays unchanged. Once a capability has been demonstrated, the cost of reproducing it tends to fall. Each drop in cost pulls new actors past ...