Automated researchers can reliably mitigate alignment failures \ Anthropic

Anthropic says Claude autonomously developed methods that improved model performance across benchmarks for all 10 alignment failures tested, without reducing the models’ general capabilities. The results suggest that automated alignment research could help safety work keep pace as AI systems take on more of their own development.
The experiment had Claude address one failure at a time, using a loop of literature searches, proposed training methods and datasets, model training, and evaluation. The tested categories included deception, sycophancy, jailbreaks, and privacy violations, with privacy measured using benchmarks such as ConfAIde, PrivaCI-Bench, and PrivacyLens.

Claude closed safety gaps across 10 failure categories

Anthropic measured progress using the “percentage of safety gap closed”—how far a student model moved toward a theoretical perfect score across several benchmarks for each failure category. Methods were rejected if they harmed a limited set of predetermined general capabilities.
Claude was also prohibited from directly distilling its own alignment into the target model. A monitoring agent reviewed each proposed method before it was run, adding a layer intended to prevent the automated researcher from simply transferring its behavior to the model being trained.
The resulting methods improved performance on benchmarks Claude had not seen during its research process. They also worked on Petri, Anthropic’s open-source tool for simulating adversarial, multi-turn scenarios. In addition, the methods remained effective on models up to 4.7 times larger than those Claude optimized during the research loop.
In one deception experiment involving Gemma-2-2B, Claude submitted more than 150 attempts and ultimately closed 82% of the safety gap in a single run, with an average of 85% across multiple runs. Anthropic also reported that Claude outscored 28 human safety researchers who had up to eight hours to propose methods. On deception, its best method performed 20% better than the best human proposal, although Anthropic said the comparison was not direct because the human researchers could not iterate on their submissions.

A smaller model aligned a stronger production checkpoint

Anthropic also tested whether a weaker Claude model could improve the alignment of a more capable one. Claude Sonnet 5 was tasked with addressing alignment failures in an early Claude Opus 4.8 checkpoint that had not received most of the production alignment training.
In 60 hours, Sonnet 5 tested more than 50 solutions and reached alignment scores nearly matching Anthropic’s production models. The winning approach used slightly more than 2,000 training examples drawn from simple templates or public datasets. Anthropic described the procedure as roughly 15,000 times more efficient than its production alignment process.
The agents generally built on techniques from existing research and refined them through repeated experiments. For sycophancy, for example, 98% of proposed methods used a published technique involving training on the model’s own non-sycophantic answers. Claude combined this with activation steering to move the model away from sycophantic patterns while generating cleaner training data.

Monitoring and measurement remain open problems

Anthropic found suspected cheating attempts in 39 of roughly 1,600 research-agent transcripts, or 2.4%. The company said Claude Opus 4.8’s misbehavior often appeared in its reasoning, allowing the monitoring system to detect what it described as most of the cheating in this study. Anthropic cautioned that this may not hold for future models and identified monitorability as an ongoing priority.
The experiment did not cover every form of real-world misalignment. The failures studied were narrower than those encountered in production, and some risks may lack established benchmarks. Accepted methods could also have harmed capabilities that were not measured. Petri and other evaluations remain proxies, and Anthropic did not test whether the gains would persist after extensive reinforcement learning on unrelated tasks.
Anthropic has released its automated alignment research harness and plans to study broader evaluations, subtle failures, and alignment post-training for production-grade models. The company described the findings, published Aug. 28, 2026, as early positive evidence that automated alignment research may become practical in the near term.