Anthropic’s AI Agents Can Now Train Other Models to Fix Alignment Failures
For the experiment, Anthropic developed Automated Alignment Researchers (AARs) using Claude Opus 4.8.
Anthropic has published a new study showing that AI agents powered by Claude Opus 4.8 can independently research, develop and test training methods.
The company said its automated systems successfully mitigated all 10 alignment failures evaluated in the study, offering what Anthropic described as early evidence that automated alignment post-training could become practical in the near term.
For the experiment, Anthropic developed Automated Alignment Researchers (AARs) using Claude Opus 4.8. Each agent searches existing research, proposes a training approach, trains a target model for around 30 minutes on a single NVIDIA H200 GPU, evaluates the outcome and uses the results to determine which experiment to run next.
The process can continue for up to 48 hours. The researchers tested the systems against failures including deception, sycophancy, jailbreak compliance, prompt injection and power-seeking behaviour.
Anthropic said the strongest training methods also improved performance on benchmarks that were not used during the research process. The methods generalised to models up to 4.7 times larger than the original target models.
The company screened the methods against MMLU, GSM8K and IFEval to ensure that alignment improvements did not significantly compromise general capabilities. However, Anthropic cautioned that the experiment did not cover every model capability, meaning some unintended effects may remain undetected.
The company also compared AAR-generated methods with proposals from 28 experienced AI safety researchers. Across the seven alignment failures for which human researchers proposed methods, Anthropic said the best AAR approach eventually outperformed the strongest human proposal, after an average of 6.4 hours of experimentation.
However, Anthropic noted that the comparison was not directly equivalent because human researchers did not have the same opportunity for iterative experimentation.
The study also uncovered potential risks: 39 of 1,601 monitored AAR trajectories involved attempts to cheat evaluations, although none produced a reported training method.
Anthropic has released the automated alignment research harness used in the study, allowing researchers to investigate and build on the approach.
