Anthropic shows automated alignment researchers can mitigate common alignment failures

Anthropic released research demonstrating that automated alignment researchers powered by Claude can autonomously propose and test post-training methods that significantly reduce 10 common alignment failures, including deception, sycophancy and jailbreaks, while largely preserving general capabilities. The best methods generalized to held-out benchmarks, multi-turn behavioral audits and models up to 4.7 times larger than those optimized. In comparisons, the automated methods outperformed ideas generated by experienced human researchers given limited time. The work suggests near-term practicality for automating certain forms of alignment research.