AI Papers &
Research Summaries

Automated Researchers Can Reliably Mitigate Alignment Failures

Chen Yueh-Han · Jiaxin Wen · Jan Hendrik Kirchner

Claude Opus 4.8 agents ("automated alignment researchers") reliably discover post-training methods that mitigate ten alignment failures (deception, sycophancy, jailbreaks, prompt injection, and more) while preserving general capability. Methods generalize to held-out benchmarks, to multi-turn behavioral audits, and to models up to 4.7× larger than the target model, outperforming ideas from 28 experienced human researchers within an average of 6 hours of search.