Anthropic’s AI Researchers Automate Alignment Gains

AI alignment research aims to curb behaviors such as deception, sycophancy and compliance with jailbreaks, but progress is constrained by the time researchers need to design, train and test each intervention. Anthropic’s Automated Alignment Researcher, or AAR, uses Claude agents to search literature, propose methods, conduct post-training and evaluate results in repeated loops. The work shows AI can take over parts of measurable safety research while preserving broader capabilities, offering early evidence for automated model research and, eventually, recursive self-improvement.

Anthropic published “Automated Researchers Can Reliably Mitigate Alignment Failures” on Aug. 28, 2026. The system improved results across benchmarks covering 10 alignment failures and generalized to a held-out test, multi-turn behavioral audits and models up to 4.7 times larger. Its best methods surpassed ideas from 28 experienced researchers after an average of six hours, while human guidance did not improve the agents’ results. Anthropic said an AAR costs about $4 an hour in API inference, compared with roughly $150 an hour for human researchers.