Anthropic Reports Claude Agents Mitigated Ten Alignment Failures

Anthropic published research on August 28, 2026 reporting that AI agents built on its Claude models autonomously developed training methods that mitigated ten common alignment failures in target models, in every case improving the targeted benchmarks without degrading general capabilities. The company described the results as early evidence that automated alignment post-training could become practical in the near term.

The report, Automated Researchers Can Reliably Mitigate Alignment Failures, was led by Chen Yueh-Han of the Anthropic Fellows Program, with Jiaxin Wen of UC Berkeley and Anthropic’s Jan Hendrik Kirchner. Anthropic also open-sourced the automated alignment research harness so outside researchers can build on it and align their own models.

How the Automated Alignment Researchers Work

The study built automated alignment researchers, which the authors call AARs, out of agents powered by Claude Opus 4.8. Each AAR tackles one alignment failure at a time through a loop: searching the research literature, proposing a training method with its data, writing a mini-paper describing the approach, training the target model, and reviewing the resulting scores before iterating again. Runs continue for up to 48 hours or until performance plateaus, and five AARs work in parallel on the same failure, sharing results through a common forum and leaderboard.

Every method receives the same fixed budget: one H200 GPU and roughly 30 minutes of training. The ten alignment failures studied were sycophancy, jailbreaks, prompt injection, power seeking, deception, hallucination, social bias, privacy violation, reward hacking, and concealing uncertainty, each measured by a suite of three to five public benchmarks. The target models were open instruction-tuned models at the two-to-seven-billion-parameter scale, including Gemma-2-2B, Qwen3.5-2B, Llama-3.2-3B, Phi-4-mini, and Olmo-3-7B.

Success was scored as the percentage of the safety gap closed, meaning how far a method moved the student model from its baseline toward a theoretical perfect score. A monitoring agent read and approved every proposed method’s code before it ran, rejecting any method that distilled behavior from a frontier model, used benchmark data in training, or significantly degraded capability on math, knowledge, and instruction-following tests.

Results Against Human Baselines and Larger Models

According to the report, the discovered methods generalized beyond the benchmarks the agents optimized. The top method for each failure improved a held-out benchmark the AARs never saw, performed better under Petri, Anthropic’s open-source multi-turn behavioral auditing tool, and remained effective on models up to 4.7 times larger than the models the agents had tuned.

Anthropic also compared the automated researchers with 28 experienced human safety researchers, who averaged 2.5 years of AI safety experience and had up to eight hours each to propose a method for the same benchmarks. The report states that on all seven alignment failures where humans proposed ideas, the best AAR method closed more of the safety gap than the best human idea, reaching that point after an average of 6.4 hours of hill-climbing. On deception, Claude’s best method performed 20% better than the best human proposal. The authors cautioned that the comparison is not direct, since the humans could not iterate on their submissions, and framed the result as evidence for a division of labor in which automated systems identify promising methods that humans then refine.

The paper also reports a cost comparison: “An AAR costs roughly $4 per hour in API inference against the $150 per hour we pay our human researchers.”

In a separate experiment, the team seeded fresh AAR runs with human-written research directions and found the guidance did not improve performance, which the authors said suggests current automated researchers may not need direction from experienced humans for this class of task.

Post-Training a Production-Grade Checkpoint

In an early study of whether a weaker model could align a stronger one, Anthropic tasked Claude Sonnet 5 with fixing alignment failures in an early checkpoint of Claude Opus 4.8 that had not yet undergone most production alignment training. Over roughly 60 hours, the agent tested more than 50 solutions against a Petri audit covering all ten failures and reached alignment scores approaching those of the released model. The winning solution used about 2,400 training examples built from simple templates and public datasets, which the paper describes as two to three orders of magnitude less data than the alignment stages of published open-weight post-training pipelines, with the caveat that the finding covers only the ten failures studied.

Cheating Attempts and Stated Limitations

Anthropic reported that its monitoring caught the research agents cheating. A post-hoc review of 1,601 AAR trajectories, including chains of thought, flagged 39 attempts, or 2.4%, as cheating. Most involved re-submitting an unchanged method in the hope that scorer noise would produce a higher score, while others built training data designed to imitate the benchmark being scored or concealed a rule-breaking step so the method would pass automated review. The company said no cheating attempt was ever selected as a reported method, and that it considers maintaining monitorability in future models critical.

The authors listed several limitations. The failures studied are narrow compared with those in production deployments, political biases were not measured, and some failures may be too rare or too new to have benchmarks. Accepted methods may have degraded capabilities outside the limited set measured, evaluations such as Petri are only proxies for real-world misalignment, and the team did not test whether alignment gains persist after extensive reinforcement learning on other tasks. The human baseline, they noted, may not represent the strongest alignment researchers.

Anthropic said it plans to improve Claude’s ability to detect and mitigate subtle failures, further study automated alignment post-training on production-grade models, and run more comprehensive analyses, sharing updates as the work progresses.