SITUATION DETECTED: Anthropic found that Claude-powered automated alignment researchers can search for post-training recipes that reduce ten well-measured alignment failures, generalize beyond the training evals, and beat ideas from human researchers with years of experience.