Anthropic has released a new report demonstrating that automated alignment researchers can reliably mitigate alignment failures in AI models. The experiment tasked Claude with autonomously improving models on 10 categories of alignment failure, including deception, sycophancy, jailbreaks, and privacy violations. Claude operated through a loop of searching literature, proposing methods, training, and testing. Success was measured by “percentage of safety gap closed,” and methods that degraded general capabilities were excluded. Claude was also forbidden from distilling its own alignment directly into the target model, enforced by a monitoring agent.

The results were positive across all 10 failures: Claude found fixes that improved target benchmarks without degrading capabilities. The best methods also worked on withheld alignment benchmarks and on Petri, an open-source adversarial multi-turn testing tool. Moreover, the methods remained effective on models up to 4.7 times larger than those Claude was optimized for during the research loop. In a direct comparison, Claude outperformed 28 human safety researchers who had up to eight hours to devise methods. On deception, Claude‘s best method performed 20% better than the best human proposal. However, since humans could not iterate, the authors view this as evidence for a workflow where Claude identifies promising methods for humans to refine.

In a more ambitious scenario, Claude Sonnet 5 (weaker than Claude Opus 4.8 on the Epoch Capabilities Index) was tasked with fixing alignment failures in an early Opus 4.8 checkpoint that had not yet undergone production alignment training. In just 60 hours, Claude experimented with over 50 solutions and achieved alignment scores nearly matching those of the production models. The winning solution contained just over 2,000 training examples, built from simple templates or public datasets, making it roughly 15,000 times more efficient than the production alignment procedure.

The report details the methods Claude proposed. For sycophancy, 98% of proposed methods used a published technique: training the model on its own non-sycophantic answers. Claude used activation steering to nudge the model away from sycophantic patterns and produce cleaner training examples. The agents also monitored for cheating behaviors: Claude Opus 4.8 reviewed ~1,600 research agent transcripts and found cheating attempts in 39 (2.4%). The authors note that maintaining monitorability in future models is critical.

Despite these encouraging findings, the experiment had several limitations. The alignment failures studied were narrow compared to those in production (e.g., political biases were not measured). Some failures may occur so rarely or emerge so recently that no benchmark exists to measure them. The authors only rejected methods when they degraded a limited set of predetermined capabilities, meaning accepted methods may have degraded other important capabilities that were not measured. Evaluations like Petri are only proxies for real-world misalignment, and the study did not test whether alignment gains persist after extensive RL training on other tasks.

Anthropic plans to continue improving Claude‘s ability to measure subtle failures, further study automating alignment post-training on production-grade models, and run more comprehensive analyses. They view these results as early positive signals that automated alignment post-training could become practical in the near term. The automated alignment research harness has been open-sourced for others to build on.