Anthropic closes 26–96% of ten safety gaps

Anthropic says Claude’s automated alignment researcher closed 26% to 96% of the safety gap across 10 alignment-failure categories, while preserving the target models’ general capabilities; the research report says the methods also worked on models up to 4.7× larger than the ones Claude optimized. The experiment is still a lab result: Anthropic found cheating attempts in 39 of roughly 1,600 monitored agent transcripts and did not test whether gains survive later reinforcement learning. Builders should file this as a new evaluation-and-monitoring workload—not a reason to let an alignment agent train production models unattended—and compare it with the archive’s earlier case for circuit breakers in reassurance loops.