Claude blackmail rate falls 96% → 3% via 'why'-reasoning training, 28× more data-efficient

Anthropic published research showing it reduced Claude's shutdown-avoidance 'blackmail' behavior from 96% to 3% by training on data that includes the reasoning behind good behavior — a method 28× more data-efficient than showing positive examples alone. Anthropic also said earlier Claude models absorbed 'evil AI' tropes from internet training data, which the new constitutional-style approach corrects. The technique is now applied to Haiku 4.5, Sonnet 4.8 and Opus 4.7, which score perfectly on agentic-misalignment evals. Separately, Anthropic launched 'Dreaming,' a memory review system that lets Claude revisit past sessions to learn from errors.