Anthropic releases automated alignment platform as agents outpace experts

Anthropic announced today the infrastructure it built for Automated Alignment Research, AAR, a suite of agents based on Claude Opus 4.8 that scan literature, propose post-training methods and execute them on a target model for about half an hour on a single H200 GPU. The central restriction is that the methods may not distill behavior from the AAR itself or from a stronger model, so any improvement must stem from the method alone. Any candidate that harms general capability measured on MMLU, GSM8K or IFEval is automatically disqualified.

The experiment covered ten well-known alignment failures—deception, sycophancy, jailbreak compliance and others—each already represented in public benchmarks such as MASK for deception and HarmBench for jailbreaks. The strongest methods generated by the AARs markedly reduced the targeted failures, and the gains generalized to a held-out benchmark, multi-round behavioral evaluations (Petri), and models up to 4.7 times larger than the model on which they were trained.

For a human baseline, 28 experienced researchers were given up to eight hours to develop methods for the same benchmarks. Their approaches performed worse than the leading AARs. When human ideas were used only as an initial direction for the agents, performance did not improve, suggesting that current AARs do not require human guidance to discover effective solutions.

The choice of post-alignment as a testbed is intentional. Unlike scalable oversight or knowledge extraction tasks, this setting provides an objective metric—the benchmark—that determines whether a fix works without relying on human judgment that might miss errors. Bowkis et al. (2026) argue that automating a researcher can be hazardous when no clear external metric exists; here the metric is present, and the loop of proposal, execution, and generalization test closes rapidly.

The full report, authored by Chen Yueh-Han, Jiaxin Wen and Jan Hendrik Kirchner as part of Anthropic’s fellows program, concludes that automating alignment research on well-characterized failures may become practical in the near term. Anthropic is now releasing the automated setup so other groups can build upon it, turning the loop into a shared resource rather than a single-lab artifact.