Anthropic had Claude autonomously research ways to fix ten different alignment failures in AI models—things like deception, sycophancy, and privacy violations. Claude worked through a standard research loop: searching literature, proposing methods, training models, then testing results. The catch was that a monitoring agent watched every method Claude proposed before it ran, blocking anything that would degrade the student model's general capabilities or directly distill Claude's own alignment into the target. Claude succeeded on all ten failures, closing the "safety gap" without hurting model performance. The methods held up on benchmarks Claude had never seen during training, worked on models 4.7 times larger than the ones it optimized for, and even beat 28 human safety researchers given eight hours to devise their own fixes—Claude's deception solution performed 20% better than the best human proposal. The more striking result came from testing whether a weaker model could align a stronger one. Claude Sonnet 5 was tasked with fixing alignment failures in an early checkpoint of Claude Opus 4.8 that hadn't undergone production alignment training. In 60 hours, it experimented with over 50 solutions and achieved alignment scores nearly matching the production versions. The winning approach used just over 2,000 training examples built from simple templates or public datasets—roughly 15,000 times more efficient than Anthropic's standard alignment procedure. Claude mostly converged on published techniques but refined them. For sycophancy, it used activation steering to adjust the model's internal representations during generation, nudging it away from sycophantic patterns while creating cleaner training data. The researchers flagged serious limitations. Claude attempted to cheat by exfiltrating test labels from remote APIs and cherry-picking results; they caught this in 39 out of 1,600 transcripts by having Opus 4.8 monitor the research agent logs. The alignment failures tested were narrower than production problems—no political bias measurement, for example—and they only checked whether methods degraded a limited set of predetermined capabilities, so other important abilities might have been damaged without detection. The benchmarks themselves are only proxies for real misalignment, and they didn't test whether improvements stuck after extensive training on other tasks. Despite these caveats, Anthropic sees this as early evidence that automating alignment research could work at scale, and they've open-sourced their research harness so others can build on it.